Search

Diffusion models are successful for synthesizing high-quality videos but are limited to generating short clips (e.g., 2-10 seconds). Synthesizing sustained footage (e.g. over minutes) still remains an open research question. In this paper, we propose MALT Diffusion (using Memory-Augmented Latent Transformers), a new diffusion model specialized for long video generation. MALT Diffusion (or just MALT) handles long videos by subdividing them into short segments and doing segment-level autoregressive generation. To achieve this, we first propose recurrent attention layers that encode multiple segments into a compact memory latent vector; by maintaining this memory vector over time, MALT is able to condition on it and continuously generate new footage based on a long temporal context. We also present several training techniques that enable the model to generate frames over a long horizon with consistent quality and minimal degradation. We validate the effectiveness of MALT through experiments on long video benchmarks. We first perform extensive analysis of MALT in long-contextual understanding capability and stability using popular long video benchmarks. For example, MALT achieves an FVD score of 220.4 on 128-frame video generation on UCF-101, outperforming the previous state-of-the-art of 648.4. Finally, we explore MALT's capabilities in a text-to-video generation setting and show that it can produce long videos compared with recent techniques for long text-to-video generation., Comment: preprint. 26 pages

Aligning text-to-image (T2I) diffusion models with preference optimization is valuable for human-annotated datasets, but the heavy cost of manual data collection limits scalability. Using reward models offers an alternative, however, current preference optimization methods fall short in exploiting the rich information, as they only consider pairwise preference distribution. Furthermore, they lack generalization to multi-preference scenarios and struggle to handle inconsistencies between rewards. To address this, we present Calibrated Preference Optimization (CaPO), a novel method to align T2I diffusion models by incorporating the general preference from multiple reward models without human annotated data. The core of our approach involves a reward calibration method to approximate the general preference by computing the expected win-rate against the samples generated by the pretrained models. Additionally, we propose a frontier-based pair selection method that effectively manages the multi-preference distribution by selecting pairs from Pareto frontiers. Finally, we use regression loss to fine-tune diffusion models to match the difference between calibrated rewards of a selected pair. Experimental results show that CaPO consistently outperforms prior methods, such as Direct Preference Optimization (DPO), in both single and multi-reward settings validated by evaluation on T2I benchmarks, including GenEval and T2I-Compbench.

In this paper we focus on inserting a given human (specifically, a single image of a person) into a novel scene. Our method, which builds on top of Stable Diffusion, yields natural looking images while being highly controllable with text and pose. To accomplish this we need to train on pairs of images, the first a reference image with the person, the second a "target image" showing the same person (with a different pose and possibly in a different background). Additionally we require a text caption describing the new pose relative to that in the reference image. In this paper we present a novel dataset following this criteria, which we create using pairs of frames from human-centric and action-rich videos and employing a multimodal LLM to automatically summarize the difference in human pose for the text captions. We demonstrate that identity preservation is a more challenging task in scenes "in-the-wild", and especially scenes where there is an interaction between persons and objects. Combining the weak supervision from noisy captions, with robust 2D pose improves the quality of person-object interactions.

Recent advancements in large language model(LLM) performance on medical multiple choice question (MCQ) benchmarks have stimulated interest from healthcare providers and patients globally. Particularly in low-and middle-income countries (LMICs) facing acute physician shortages and lack of specialists, LLMs offer a potentially scalable pathway to enhance healthcare access and reduce costs. However, their effectiveness in the Global South, especially across the African continent, remains to be established. In this work, we introduce AfriMed-QA, the first large scale Pan-African English multi-specialty medical Question-Answering (QA) dataset, 15,000 questions (open and closed-ended) sourced from over 60 medical schools across 16 countries, covering 32 medical specialties. We further evaluate 30 LLMs across multiple axes including correctness and demographic bias. Our findings show significant performance variation across specialties and geographies, MCQ performance clearly lags USMLE (MedQA). We find that biomedical LLMs underperform general models and smaller edge-friendly LLMs struggle to achieve a passing score. Interestingly, human evaluations show a consistent consumer preference for LLM answers and explanations when compared with clinician answers.

Foundational models are able to generate text outputs given prompt instructions and text, audio, or image inputs. Recently these models have been combined to perform tasks on video, such as video summarization. Such video foundation models perform pre-training by aligning outputs from each modality-specific model into the same embedding space. Then the embeddings from each model are used within a language model, which is fine-tuned on a desired instruction set. Aligning each modality during pre-training is computationally expensive and prevents rapid testing of different base modality models. During fine-tuning, evaluation is carried out within in-domain videos where it is hard to understand the generalizability and data efficiency of these methods. To alleviate these issues we propose a plug-and-play video language model. It directly uses the texts generated from each input modality into the language model, avoiding pre-training alignment overhead. Instead of fine-tuning we leverage few-shot instruction adaptation strategies. We compare the performance versus the computational costs for our plug-and-play style method and baseline tuning methods. Finally, we explore the generalizability of each method during domain shift and present insights on what data is useful when training data is limited. Through this analysis, we present practical insights on how to leverage multi-modal foundational models for effective results given realistic compute and data limitations., Comment: 11 pages, 4 figures

Video Language Models (VLMs) are crucial for generalizing across diverse tasks and using language cues to enhance learning. While transformer-based architectures have been the de facto in vision-language training, they face challenges like quadratic computational complexity, high GPU memory usage, and difficulty with long-term dependencies. To address these limitations, we introduce MambaVL, a novel model that leverages recent advancements in selective state space modality fusion to efficiently capture long-range dependencies and learn joint representations for vision and language data. MambaVL utilizes a shared state transition matrix across both modalities, allowing the model to capture information about actions from multiple perspectives within the scene. Furthermore, we propose a question-answering task that helps guide the model toward relevant cues. These questions provide critical information about actions, objects, and environmental context, leading to enhanced performance. As a result, MambaVL achieves state-of-the-art performance in action recognition on the Epic-Kitchens-100 dataset and outperforms baseline methods in action anticipation.

Cross-modal contrastive pre-training between natural language and other modalities, e.g., vision and audio, has demonstrated astonishing performance and effectiveness across a diverse variety of tasks and domains. In this paper, we investigate whether such natural language supervision can be used for wearable sensor based Human Activity Recognition (HAR), and discover that-surprisingly-it performs substantially worse than standard end-to-end training and self-supervision. We identify the primary causes for this as: sensor heterogeneity and the lack of rich, diverse text descriptions of activities. To mitigate their impact, we also develop strategies and assess their effectiveness through an extensive experimental evaluation. These strategies lead to significant increases in activity recognition, bringing performance closer to supervised and self-supervised training, while also enabling the recognition of unseen activities and cross modal retrieval of videos. Overall, our work paves the way for better sensor-language learning, ultimately leading to the development of foundational models for HAR using wearables.

The goal of image cropping is to identify visually appealing crops within an image. Conventional methods rely on specialized architectures trained on specific datasets, which struggle to be adapted to new requirements. Recent breakthroughs in large vision-language models (VLMs) have enabled visual in-context learning without explicit training. However, effective strategies for vision downstream tasks with VLMs remain largely unclear and underexplored. In this paper, we propose an effective approach to leverage VLMs for better image cropping. First, we propose an efficient prompt retrieval mechanism for image cropping to automate the selection of in-context examples. Second, we introduce an iterative refinement strategy to iteratively enhance the predicted crops. The proposed framework, named Cropper, is applicable to a wide range of cropping tasks, including free-form cropping, subject-aware cropping, and aspect ratio-aware cropping. Extensive experiments and a user study demonstrate that Cropper significantly outperforms state-of-the-art methods across several benchmarks.

We extend multimodal transformers to include 3D camera motion as a conditioning signal for the task of video generation. Generative video models are becoming increasingly powerful, thus focusing research efforts on methods of controlling the output of such models. We propose to add virtual 3D camera controls to generative video methods by conditioning generated video on an encoding of three-dimensional camera movement over the course of the generated video. Results demonstrate that we are (1) able to successfully control the camera during video generation, starting from a single frame and a camera signal, and (2) we demonstrate the accuracy of the generated 3D camera paths using traditional computer vision methods.

We present SLAIM - Simultaneous Localization and Implicit Mapping. We propose a novel coarse-to-fine tracking model tailored for Neural Radiance Field SLAM (NeRF-SLAM) to achieve state-of-the-art tracking performance. Notably, existing NeRF-SLAM systems consistently exhibit inferior tracking performance compared to traditional SLAM algorithms. NeRF-SLAM methods solve camera tracking via image alignment and photometric bundle-adjustment. Such optimization processes are difficult to optimize due to the narrow basin of attraction of the optimization loss in image space (local minima) and the lack of initial correspondences. We mitigate these limitations by implementing a Gaussian pyramid filter on top of NeRF, facilitating a coarse-to-fine tracking optimization strategy. Furthermore, NeRF systems encounter challenges in converging to the right geometry with limited input views. While prior approaches use a Signed-Distance Function (SDF)-based NeRF and directly supervise SDF values by approximating ground truth SDF through depth measurements, this often results in suboptimal geometry. In contrast, our method employs a volume density representation and introduces a novel KL regularizer on the ray termination distribution, constraining scene geometry to consist of empty space and opaque surfaces. Our solution implements both local and global bundle-adjustment to produce a robust (coarse-to-fine) and accurate (KL regularizer) SLAM solution. We conduct experiments on multiple datasets (ScanNet, TUM, Replica) showing state-of-the-art results in tracking and in reconstruction accuracy.

We study the task of 3D multi-object re-identification from embodied tours. Specifically, an agent is given two tours of an environment (e.g. an apartment) under two different layouts (e.g. arrangements of furniture). Its task is to detect and re-identify objects in 3D - e.g. a "sofa" moved from location A to B, a new "chair" in the second layout at location C, or a "lamp" from location D in the first layout missing in the second. To support this task, we create an automated infrastructure to generate paired egocentric tours of initial/modified layouts in the Habitat simulator using Matterport3D scenes, YCB and Google-scanned objects. We present 3D Semantic MapNet (3D-SMNet) - a two-stage re-identification model consisting of (1) a 3D object detector that operates on RGB-D videos with known pose, and (2) a differentiable object matching module that solves correspondence estimation between two sets of 3D bounding boxes. Overall, 3D-SMNet builds object-based maps of each layout and then uses a differentiable matcher to re-identify objects across the tours. After training 3D-SMNet on our generated episodes, we demonstrate zero-shot transfer to real-world rearrangement scenarios by instantiating our task in Replica, Active Vision, and RIO environments depicting rearrangements. On all datasets, we find 3D-SMNet outperforms competitive baselines. Further, we show jointly training on real and generated episodes can lead to significant improvements over training on real data alone., Comment: 8pages

Anticipating future actions is a highly challenging task due to the diversity and scale of potential future actions; yet, information from different modalities help narrow down plausible action choices. Each modality can provide diverse and often complementary context for the model to learn from. While previous multi-modal methods leverage information from modalities such as video and audio, we primarily explore how text descriptions of actions and objects can also lead to more accurate action anticipation by providing additional contextual cues, e.g., about the environment and its contents. We propose a Multi-modal Contrastive Anticipative Transformer (M-CAT), a video transformer architecture that jointly learns from multi-modal features and text descriptions of actions and objects. We train our model in two stages, where the model first learns to align video clips with descriptions of future actions, and is subsequently fine-tuned to predict future actions. Compared to existing methods, M-CAT has the advantage of learning additional context from two types of text inputs: rich descriptions of future actions during pre-training, and, text descriptions for detected objects and actions during modality feature fusion. Through extensive experimental evaluation, we demonstrate that our model outperforms previous methods on the EpicKitchens datasets, and show that using simple text descriptions of actions and objects aid in more effective action anticipation. In addition, we examine the impact of object and action information obtained via text, and perform extensive ablations.

Recent works have demonstrated that using reinforcement learning (RL) with multiple quality rewards can improve the quality of generated images in text-to-image (T2I) generation. However, manually adjusting reward weights poses challenges and may cause over-optimization in certain metrics. To solve this, we propose Parrot, which addresses the issue through multi-objective optimization and introduces an effective multi-reward optimization strategy to approximate Pareto optimal. Utilizing batch-wise Pareto optimal selection, Parrot automatically identifies the optimal trade-off among different rewards. We use the novel multi-reward optimization algorithm to jointly optimize the T2I model and a prompt expansion network, resulting in significant improvement of image quality and also allow to control the trade-off of different rewards using a reward related prompt during inference. Furthermore, we introduce original prompt-centered guidance at inference time, ensuring fidelity to user input after prompt expansion. Extensive experiments and a user study validate the superiority of Parrot over several baselines across various quality criteria, including aesthetics, human preference, text-image alignment, and image sentiment.

We present VideoPoet, a language model capable of synthesizing high-quality video, with matching audio, from a large variety of conditioning signals. VideoPoet employs a decoder-only transformer architecture that processes multimodal inputs -- including images, videos, text, and audio. The training protocol follows that of Large Language Models (LLMs), consisting of two stages: pretraining and task-specific adaptation. During pretraining, VideoPoet incorporates a mixture of multimodal generative objectives within an autoregressive Transformer framework. The pretrained LLM serves as a foundation that can be adapted for a range of video generation tasks. We present empirical results demonstrating the model's state-of-the-art capabilities in zero-shot video generation, specifically highlighting VideoPoet's ability to generate high-fidelity motions. Project page: http://sites.research.google/videopoet/, Comment: To appear at ICML 2024; Project page: http://sites.research.google/videopoet/

We present W.A.L.T, a transformer-based approach for photorealistic video generation via diffusion modeling. Our approach has two key design decisions. First, we use a causal encoder to jointly compress images and videos within a unified latent space, enabling training and generation across modalities. Second, for memory and training efficiency, we use a window attention architecture tailored for joint spatial and spatiotemporal generative modeling. Taken together these design decisions enable us to achieve state-of-the-art performance on established video (UCF-101 and Kinetics-600) and image (ImageNet) generation benchmarks without using classifier free guidance. Finally, we also train a cascade of three models for the task of text-to-video generation consisting of a base latent video diffusion model, and two video super-resolution diffusion models to generate videos of $512 \times 896$ resolution at $8$ frames per second., Comment: Project website https://walt-video-diffusion.github.io/

Domain randomization (DR), which entails training a policy with randomized dynamics, has proven to be a simple yet effective algorithm for reducing the gap between simulation and the real world. However, DR often requires careful tuning of randomization parameters. Methods like Bayesian Domain Randomization (Bayesian DR) and Active Domain Randomization (Adaptive DR) address this issue by automating parameter range selection using real-world experience. While effective, these algorithms often require long computation time, as a new policy is trained from scratch every iteration. In this work, we propose Adaptive Bayesian Domain Randomization via Strategic Fine-tuning (BayRnTune), which inherits the spirit of BayRn but aims to significantly accelerate the learning processes by fine-tuning from previously learned policy. This idea leads to a critical question: which previous policy should we use as a prior during fine-tuning? We investigated four different fine-tuning strategies and compared them against baseline algorithms in five simulated environments, ranging from simple benchmark tasks to more complex legged robot environments. Our analysis demonstrates that our method yields better rewards in the same amount of timesteps compared to vanilla domain randomization or Bayesian DR.

While Large Language Models (LLMs) are the dominant models for generative tasks in language, they do not perform as well as diffusion models on image and video generation. To effectively use LLMs for visual generation, one crucial component is the visual tokenizer that maps pixel-space inputs to discrete tokens appropriate for LLM learning. In this paper, we introduce MAGVIT-v2, a video tokenizer designed to generate concise and expressive tokens for both videos and images using a common token vocabulary. Equipped with this new tokenizer, we show that LLMs outperform diffusion models on standard image and video generation benchmarks including ImageNet and Kinetics. In addition, we demonstrate that our tokenizer surpasses the previously top-performing video tokenizer on two more tasks: (1) video compression comparable to the next-generation video codec (VCC) according to human evaluations, and (2) learning effective representations for action recognition tasks., Comment: ICLR 2024

Humanoid robots are well suited for human habitats due to their morphological similarity, but developing controllers for them is a challenging task that involves multiple sub-problems, such as control, planning and perception. In this paper, we introduce a method to simplify controller design by enabling users to train and fine-tune robot control policies using natural language commands. We first learn a neural network policy that generates behaviors given a natural language command, such as "walk forward", by combining Large Language Models (LLMs), motion retargeting, and motion imitation. Based on the synthesized motion, we iteratively fine-tune by updating the text prompt and querying LLMs to find the best checkpoint associated with the closest motion in history. We validate our approach using a simulated Digit humanoid robot and demonstrate learning of diverse motions, such as walking, hopping, and kicking, without the burden of complex reward engineering. In addition, we show that our iterative refinement enables us to learn 3x times faster than a naive formulation that learns from scratch.

Web articles such as Wikipedia serve as one of the major sources of knowledge dissemination and online learning. However, their in-depth information--often in a dense text format--may not be suitable for mobile browsing, even in a responsive UI. We propose an automatic approach that converts a structural article of any length into a set of interactive Web Stories that are ideal for mobile experiences. We focused on Wikipedia articles and developed Wiki2Story, a pipeline based on language and layout models, to demonstrate the concept. Wiki2Story dynamically slices an article and plans one to multiple Story paths according to the document hierarchy. For each slice, it generates a multi-page summary Story composed of text and image pairs in visually-appealing layouts. We derived design principles from an analysis of manually-created Story practices. We executed our pipeline on 500 Wikipedia documents and conducted user studies to review selected outputs. Results showed that Wiki2Story effectively captured and presented salient content from the original articles and sparked interest in viewers.

Human Activity Recognition (HAR) systems have been extensively studied by the vision and ubiquitous computing communities due to their practical applications in daily life, such as smart homes, surveillance, and health monitoring. Typically, this process is supervised in nature and the development of such systems requires access to large quantities of annotated data. However, the higher costs and challenges associated with obtaining good quality annotations have rendered the application of self-supervised methods an attractive option and contrastive learning comprises one such method. However, a major component of successful contrastive learning is the selection of good positive and negative samples. Although positive samples are directly obtainable, sampling good negative samples remain a challenge. As human activities can be recorded by several modalities like camera and IMU sensors, we propose a hard negative sampling method for multimodal HAR with a hard negative sampling loss for skeleton and IMU data pairs. We exploit hard negatives that have different labels from the anchor but are projected nearby in the latent space using an adjustable concentration parameter. Through extensive experiments on two benchmark datasets: UTD-MHAD and MMAct, we demonstrate the robustness of our approach forlearning strong feature representation for HAR tasks, and on the limited data setting. We further show that our model outperforms all other state-of-the-art methods for UTD-MHAD dataset, and self-supervised methods for MMAct: Cross session, even when uni-modal data are used during downstream activity recognition.

In this work, we introduce Semantic Pyramid AutoEncoder (SPAE) for enabling frozen LLMs to perform both understanding and generation tasks involving non-linguistic modalities such as images or videos. SPAE converts between raw pixels and interpretable lexical tokens (or words) extracted from the LLM's vocabulary. The resulting tokens capture both the semantic meaning and the fine-grained details needed for visual reconstruction, effectively translating the visual content into a language comprehensible to the LLM, and empowering it to perform a wide array of multimodal tasks. Our approach is validated through in-context learning experiments with frozen PaLM 2 and GPT 3.5 on a diverse set of image understanding and generation tasks. Our method marks the first successful attempt to enable a frozen LLM to generate image content while surpassing state-of-the-art performance in image understanding tasks, under the same setting, by over 25%., Comment: NeurIPS 2023 spotlight

We study domain-adaptive image synthesis, the problem of teaching pretrained image generative models a new style or concept from as few as one image to synthesize novel images, to better understand the compositional image synthesis. We present a framework that leverages a pretrained class-conditional generation model and visual prompt tuning. Specifically, we propose a novel source class distilled visual prompt that learns disentangled prompts of semantic (e.g., class) and domain (e.g., style) from a few images. Learned domain prompt is then used to synthesize images of any classes in the style of target domain. We conduct studies on various target domains with the number of images ranging from one to a few to many, and show qualitative results which show the compositional generalization of our method. Moreover, we show that our method can help improve zero-shot domain adaptation classification accuracy., Comment: tech report

Human activity recognition (HAR) in wearable computing is typically based on direct processing of sensor data. Sensor readings are translated into representations, either derived through dedicated preprocessing, or integrated into end-to-end learning. Independent of their origin, for the vast majority of contemporary HAR, those representations are typically continuous in nature. That has not always been the case. In the early days of HAR, discretization approaches have been explored - primarily motivated by the desire to minimize computational requirements, but also with a view on applications beyond mere recognition, such as, activity discovery, fingerprinting, or large-scale search. Those traditional discretization approaches, however, suffer from substantial loss in precision and resolution in the resulting representations with detrimental effects on downstream tasks. Times have changed and in this paper we propose a return to discretized representations. We adopt and apply recent advancements in Vector Quantization (VQ) to wearables applications, which enables us to directly learn a mapping between short spans of sensor data and a codebook of vectors, resulting in recognition performance that is generally on par with their contemporary, continuous counterparts - sometimes surpassing them. Therefore, this work presents a proof-of-concept for demonstrating how effective discrete representations can be derived, enabling applications beyond mere activity classification but also opening up the field to advanced tools for the analysis of symbolic sequences, as they are known, for example, from domains such as natural language processing. Based on an extensive experimental evaluation on a suite of wearables-based benchmark HAR tasks, we demonstrate the potential of our learned discretization scheme and discuss how discretized sensor data analysis can lead to substantial changes in HAR.

Pre-trained large text-to-image models synthesize impressive images with an appropriate use of text prompts. However, ambiguities inherent in natural language and out-of-distribution effects make it hard to synthesize image styles, that leverage a specific design pattern, texture or material. In this paper, we introduce StyleDrop, a method that enables the synthesis of images that faithfully follow a specific style using a text-to-image model. The proposed method is extremely versatile and captures nuances and details of a user-provided style, such as color schemes, shading, design patterns, and local and global effects. It efficiently learns a new style by fine-tuning very few trainable parameters (less than $1\%$ of total model parameters) and improving the quality via iterative training with either human or automated feedback. Better yet, StyleDrop is able to deliver impressive results even when the user supplies only a single image that specifies the desired style. An extensive study shows that, for the task of style tuning text-to-image models, StyleDrop implemented on Muse convincingly outperforms other methods, including DreamBooth and textual inversion on Imagen or Stable Diffusion. More results are available at our project website: https://styledrop.github.io, Comment: Preprint. Project page at https://styledrop.github.io

Text-to-image (T2I) research has grown explosively in the past year, owing to the large-scale pre-trained diffusion models and many emerging personalization and editing approaches. Yet, one pain point persists: the text prompt engineering, and searching high-quality text prompts for customized results is more art than science. Moreover, as commonly argued: "an image is worth a thousand words" - the attempt to describe a desired image with texts often ends up being ambiguous and cannot comprehensively cover delicate visual details, hence necessitating more additional controls from the visual domain. In this paper, we take a bold step forward: taking "Text" out of a pre-trained T2I diffusion model, to reduce the burdensome prompt engineering efforts for users. Our proposed framework, Prompt-Free Diffusion, relies on only visual inputs to generate new images: it takes a reference image as "context", an optional image structural conditioning, and an initial noise, with absolutely no text prompt. The core architecture behind the scene is Semantic Context Encoder (SeeCoder), substituting the commonly used CLIP-based or LLM-based text encoder. The reusability of SeeCoder also makes it a convenient drop-in component: one can also pre-train a SeeCoder in one T2I model and reuse it for another. Through extensive experiments, Prompt-Free Diffusion is experimentally found to (i) outperform prior exemplar-based image synthesis approaches; (ii) perform on par with state-of-the-art T2I models using prompts following the best practice; and (iii) be naturally extensible to other downstream applications such as anime figure generation and virtual try-on, with promising quality. Our code and models are open-sourced at https://github.com/SHI-Labs/Prompt-Free-Diffusion., Comment: Code, models and demos can be found through: https://github.com/SHI-Labs/Prompt-Free-Diffusion

Given Myanmars historical and socio-political context, hate speech spread on social media has escalated into offline unrest and violence. This paper presents findings from our remote study on the automatic detection of hate speech online in Myanmar. We argue that effectively addressing this problem will require community-based approaches that combine the knowledge of context experts with machine learning tools that can analyze the vast amount of data produced. To this end, we develop a systematic process to facilitate this collaboration covering key aspects of data collection, annotation, and model validation strategies. We highlight challenges in this area stemming from small and imbalanced datasets, the need to balance non-glamorous data work and stakeholder priorities, and closed data-sharing practices. Stemming from these findings, we discuss avenues for further work in developing and deploying hate speech detection systems for low-resource languages., Comment: ICTD 2022 Conference paper

Recent conditional image generation methods produce images of remarkable diversity, fidelity and realism. However, the majority of these methods allow conditioning only on labels or text prompts, which limits their level of control over the generation result. In this paper, we introduce MaskSketch, an image generation method that allows spatial conditioning of the generation result using a guiding sketch as an extra conditioning signal during sampling. MaskSketch utilizes a pre-trained masked generative transformer, requiring no model training or paired supervision, and works with input sketches of different levels of abstraction. We show that intermediate self-attention maps of a masked generative transformer encode important structural information of the input image, such as scene layout and object shape, and we propose a novel sampling method based on this observation to enable structure-guided generation. Our results show that MaskSketch achieves high image realism and fidelity to the guiding structure. Evaluated on standard benchmark datasets, MaskSketch outperforms state-of-the-art methods for sketch-to-image translation, as well as unpaired image-to-image translation approaches.

Animal navigation research posits that organisms build and maintain internal spatial representations, or maps, of their environment. We ask if machines -- specifically, artificial intelligence (AI) navigation agents -- also build implicit (or 'mental') maps. A positive answer to this question would (a) explain the surprising phenomenon in recent literature of ostensibly map-free neural-networks achieving strong performance, and (b) strengthen the evidence of mapping as a fundamental mechanism for navigation by intelligent embodied agents, whether they be biological or artificial. Unlike animal navigation, we can judiciously design the agent's perceptual system and control the learning paradigm to nullify alternative navigation mechanisms. Specifically, we train 'blind' agents -- with sensing limited to only egomotion and no other sensing of any kind -- to perform PointGoal navigation ('go to $\Delta$ x, $\Delta$ y') via reinforcement learning. Our agents are composed of navigation-agnostic components (fully-connected and recurrent neural networks), and our experimental setup provides no inductive bias towards mapping. Despite these harsh conditions, we find that blind agents are (1) surprisingly effective navigators in new environments (~95% success); (2) they utilize memory over long horizons (remembering ~1,000 steps of past experience in an episode); (3) this memory enables them to exhibit intelligent behavior (following walls, detecting collisions, taking shortcuts); (4) there is emergence of maps and collision detection neurons in the representations of the environment built by a blind agent as it navigates; and (5) the emergent maps are selective and task dependent (e.g. the agent 'forgets' exploratory detours). Overall, this paper presents no new techniques for the AI audience, but a surprising finding, an insight, and an explanation., Comment: Accepted to ICLR 2023

Real-world autonomous missions often require rich interaction with nearby objects, such as doors or switches, along with effective navigation. However, such complex behaviors are difficult to learn because they involve both high-level planning and low-level motor control. We present a novel framework, Cascaded Compositional Residual Learning (CCRL), which learns composite skills by recursively leveraging a library of previously learned control policies. Our framework learns multiplicative policy composition, task-specific residual actions, and synthetic goal information simultaneously while freezing the prerequisite policies. We further explicitly control the style of the motion by regularizing residual actions. We show that our framework learns joint-level control policies for a diverse set of motor skills ranging from basic locomotion to complex interactive navigation, including navigating around obstacles, pushing objects, crawling under a table, pushing a door open with its leg, and holding it open while walking through it. The proposed CCRL framework leads to policies with consistent styles and lower joint torques, which we successfully transfer to a real Unitree A1 robot without any additional fine-tuning., Comment: Website: https://www.kniranjankumar.com/ccrl/

We introduce the MAsked Generative VIdeo Transformer, MAGVIT, to tackle various video synthesis tasks with a single model. We introduce a 3D tokenizer to quantize a video into spatial-temporal visual tokens and propose an embedding method for masked video token modeling to facilitate multi-task learning. We conduct extensive experiments to demonstrate the quality, efficiency, and flexibility of MAGVIT. Our experiments show that (i) MAGVIT performs favorably against state-of-the-art approaches and establishes the best-published FVD on three video generation benchmarks, including the challenging Kinetics-600. (ii) MAGVIT outperforms existing methods in inference time by two orders of magnitude against diffusion models and by 60x against autoregressive models. (iii) A single MAGVIT model supports ten diverse generation tasks and generalizes across videos from different visual domains. The source code and trained models will be released to the public at https://magvit.cs.cmu.edu., Comment: CVPR 2023 highlight

The dichotomy between the challenging nature of obtaining annotations for activities, and the more straightforward nature of data collection from wearables, has resulted in significant interest in the development of techniques that utilize large quantities of unlabeled data for learning representations. Contrastive Predictive Coding (CPC) is one such method, learning effective representations by leveraging properties of time-series data to setup a contrastive future timestep prediction task. In this work, we propose enhancements to CPC, by systematically investigating the encoder architecture, the aggregator network, and the future timestep prediction, resulting in a fully convolutional architecture, thereby improving parallelizability. Across sensor positions and activities, our method shows substantial improvements on four of six target datasets, demonstrating its ability to empower a wide range of application scenarios. Further, in the presence of very limited labeled data, our technique significantly outperforms both supervised and self-supervised baselines, positively impacting situations where collecting only a few seconds of labeled data may be possible. This is promising, as CPC does not require specialized data transformations or reconstructions for learning effective representations.

To properly assist humans in their needs, human activity recognition (HAR) systems need the ability to fuse information from multiple modalities. Our hypothesis is that multimodal sensors, visual and non-visual tend to provide complementary information, addressing the limitations of other modalities. In this work, we propose a multi-modal framework that learns to effectively combine features from RGB Video and IMU sensors, and show its robustness for MMAct and UTD-MHAD datasets. Our model is trained in two-stage, where in the first stage, each input encoder learns to effectively extract features, and in the second stage, learns to combine these individual features. We show significant improvements of 22% and 11% compared to video only and IMU only setup on UTD-MHAD dataset, and 20% and 12% on MMAct datasets. Through extensive experimentation, we show the robustness of our model on zero shot setting, and limited annotated data setting. We further compare with state-of-the-art methods that use more input modalities and show that our method outperforms significantly on the more difficult MMact dataset, and performs comparably in UTD-MHAD dataset.

Video-based dialog task is a challenging multimodal learning task that has received increasing attention over the past few years with state-of-the-art obtaining new performance records. This progress is largely powered by the adaptation of the more powerful transformer-based language encoders. Despite this progress, existing approaches do not effectively utilize visual features to help solve tasks. Recent studies show that state-of-the-art models are biased toward textual information rather than visual cues. In order to better leverage the available visual information, this study proposes a new framework that combines 3D-CNN network and transformer-based networks into a single visual encoder to extract more robust semantic representations from videos. The visual encoder is jointly trained end-to-end with other input modalities such as text and audio. Experiments on the AVSD task show significant improvement over baselines in both generative and retrieval tasks.

We introduce a Transformer based 6D Object Pose Estimation framework VideoPose, comprising an end-to-end attention based modelling architecture, that attends to previous frames in order to estimate accurate 6D Object Poses in videos. Our approach leverages the temporal information from a video sequence for pose refinement, along with being computationally efficient and robust. Compared to existing methods, our architecture is able to capture and reason from long-range dependencies efficiently, thus iteratively refining over video sequences. Experimental evaluation on the YCB-Video dataset shows that our approach is on par with the state-of-the-art Transformer methods, and performs significantly better relative to CNN based approaches. Further, with a speed of 33 fps, it is also more efficient and therefore applicable to a variety of applications that require real-time object pose estimation. Training code and pretrained models are available at https://github.com/ApoorvaBeedu/VideoPose, Comment: arXiv admin note: text overlap with arXiv:2111.10677

We address dense action forecasting: the problem of predicting future action sequence over long durations based on partial observation. Our key insight is that future action sequences are more accurately modeled with variable, rather than one, levels of abstraction, and that the optimal level of abstraction can be dynamically selected during the prediction process. Our experiments show that most parts of future action sequences can be predicted confidently in fine detail only in small segments of future frames, which are effectively ``islands'' of high model prediction confidence in a ``sea'' of uncertainty. We propose a combination Bayesian neural network and hierarchical convolutional segmentation model to both accurately predict future actions and optimally select abstraction levels. We evaluate this approach on standard datasets against existing state-of-the-art systems and demonstrate that our ``islands of predictability'' approach maintains fine-grained action predictions while also making accurate abstract predictions where systems were previously unable to do so, and thus results in substantial, monotonic increases in accuracy.

We present Variable Experience Rollout (VER), a technique for efficiently scaling batched on-policy reinforcement learning in heterogenous environments (where different environments take vastly different times to generate rollouts) to many GPUs residing on, potentially, many machines. VER combines the strengths of and blurs the line between synchronous and asynchronous on-policy RL methods (SyncOnRL and AsyncOnRL, respectively). VER learns from on-policy experience (like SyncOnRL) and has no synchronization points (like AsyncOnRL). VER leads to significant and consistent speed-ups across a broad range of embodied navigation and mobile manipulation tasks in photorealistic 3D simulation environments. Specifically, for PointGoal navigation and ObjectGoal navigation in Habitat 1.0, VER is 60-100% faster (1.6-2x speedup) than DD-PPO, the current state of art distributed SyncOnRL, with similar sample efficiency. For mobile manipulation tasks (open fridge/cabinet, pick/place objects) in Habitat 2.0 VER is 150% faster (2.5x speedup) on 1 GPU and 170% faster (2.7x speedup) on 8 GPUs than DD-PPO. Compared to SampleFactory (the current state-of-the-art AsyncOnRL), VER matches its speed on 1 GPU, and is 70% faster (1.7x speedup) on 8 GPUs with better sample efficiency. We leverage these speed-ups to train chained skills for GeometricGoal rearrangement tasks in the Home Assistant Benchmark (HAB). We find a surprising emergence of navigation in skills that do not ostensible require any navigation. Specifically, the Pick skill involves a robot picking an object from a table. During training the robot was always spawned close to the table and never needed to navigate. However, we find that if base movement is part of the action space, the robot learns to navigate then pick an object in new environments with 50% success, demonstrating surprisingly high out-of-distribution generalization., Comment: Accepted at NeurIPS 2022

Transferring knowledge from an image synthesis model trained on a large dataset is a promising direction for learning generative image models from various domains efficiently. While previous works have studied GAN models, we present a recipe for learning vision transformers by generative knowledge transfer. We base our framework on state-of-the-art generative vision transformers that represent an image as a sequence of visual tokens to the autoregressive or non-autoregressive transformers. To adapt to a new domain, we employ prompt tuning, which prepends learnable tokens called prompt to the image token sequence, and introduce a new prompt design for our task. We study on a variety of visual domains, including visual task adaptation benchmark~\cite{zhai2019large}, with varying amount of training images, and show effectiveness of knowledge transfer and a significantly better image generation quality over existing works., Comment: technical report

Non-autoregressive generative transformers recently demonstrated impressive image generation performance, and orders of magnitude faster sampling than their autoregressive counterparts. However, optimal parallel sampling from the true joint distribution of visual tokens remains an open challenge. In this paper we introduce Token-Critic, an auxiliary model to guide the sampling of a non-autoregressive generative transformer. Given a masked-and-reconstructed real image, the Token-Critic model is trained to distinguish which visual tokens belong to the original image and which were sampled by the generative transformer. During non-autoregressive iterative sampling, Token-Critic is used to select which tokens to accept and which to reject and resample. Coupled with Token-Critic, a state-of-the-art generative transformer significantly improves its performance, and outperforms recent diffusion models and GANs in terms of the trade-off between generated image quality and diversity, in the challenging class-conditional ImageNet generation., Comment: Accepted to ECCV 2022

The emergence of self-supervised learning in the field of wearables-based human activity recognition (HAR) has opened up opportunities to tackle the most pressing challenges in the field, namely to exploit unlabeled data to derive reliable recognition systems for scenarios where only small amounts of labeled training samples can be collected. As such, self-supervision, i.e., the paradigm of 'pretrain-then-finetune' has the potential to become a strong alternative to the predominant end-to-end training approaches, let alone hand-crafted features for the classic activity recognition chain. Recently a number of contributions have been made that introduced self-supervised learning into the field of HAR, including, Multi-task self-supervision, Masked Reconstruction, CPC, and SimCLR, to name but a few. With the initial success of these methods, the time has come for a systematic inventory and analysis of the potential self-supervised learning has for the field. This paper provides exactly that. We assess the progress of self-supervised HAR research by introducing a framework that performs a multi-faceted exploration of model performance. We organize the framework into three dimensions, each containing three constituent criteria, such that each dimension captures specific aspects of performance, including the robustness to differing source and target conditions, the influence of dataset characteristics, and the feature space characteristics. We utilize this framework to assess seven state-of-the-art self-supervised methods for HAR, leading to the formulation of insights into the properties of these techniques and to establish their value towards learning representations for diverse scenarios., Comment: updated

Events across a timeline are a common data representation, seen in different temporal modalities. Individual atomic events can occur in a certain temporal ordering to compose higher level composite events. Examples of a composite event are a patient's medical symptom or a baseball player hitting a home run, caused distinct temporal orderings of patient vitals and player movements respectively. Such salient composite events are provided as labels in temporal datasets and most works optimize models to predict these composite event labels directly. We focus on uncovering the underlying atomic events and their relations that lead to the composite events within a noisy temporal data setting. We propose Neural Temporal Logic Programming (Neural TLP) which first learns implicit temporal relations between atomic events and then lifts logic rules for composite events, given only the composite events labels for supervision. This is done through efficiently searching through the combinatorial space of all temporal logic rules in an end-to-end differentiable manner. We evaluate our method on video and healthcare datasets where it outperforms the baseline methods for rule discovery., Comment: 19 pages, 5 figures

Creating visual layouts is a critical step in graphic design. Automatic generation of such layouts is essential for scalable and diverse visual designs. To advance conditional layout generation, we introduce BLT, a bidirectional layout transformer. BLT differs from previous work on transformers in adopting non-autoregressive transformers. In training, BLT learns to predict the masked attributes by attending to surrounding attributes in two directions. During inference, BLT first generates a draft layout from the input and then iteratively refines it into a high-quality layout by masking out low-confident attributes. The masks generated in both training and inference are controlled by a new hierarchical sampling policy. We verify the proposed model on six benchmarks of diverse design tasks. Experimental results demonstrate two benefits compared to the state-of-the-art layout transformer models. First, our model empowers layout transformers to fulfill controllable layout generation. Second, it achieves up to 10x speedup in generating a layout at inference time than the layout transformer baseline. Code is released at https://shawnkx.github.io/blt., Comment: ECCV 2022

We introduce a simple yet effective algorithm that uses convolutional neural networks to directly estimate object poses from videos. Our approach leverages the temporal information from a video sequence, and is computationally efficient and robust to support robotic and AR domains. Our proposed network takes a pre-trained 2D object detector as input, and aggregates visual features through a recurrent neural network to make predictions at each frame. Experimental evaluation on the YCB-Video dataset show that our approach is on par with the state-of-the-art algorithms. Further, with a speed of 30 fps, it is also more efficient than the state-of-the-art, and therefore applicable to a variety of applications that require real-time object pose estimation.

Vision Transformer (ViT) is emerging as the state-of-the-art architecture for image recognition. While recent studies suggest that ViTs are more robust than their convolutional counterparts, our experiments find that ViTs trained on ImageNet are overly reliant on local textures and fail to make adequate use of shape information. ViTs thus have difficulties generalizing to out-of-distribution, real-world data. To address this deficiency, we present a simple and effective architecture modification to ViT's input layer by adding discrete tokens produced by a vector-quantized encoder. Different from the standard continuous pixel tokens, discrete tokens are invariant under small perturbations and contain less information individually, which promote ViTs to learn global information that is invariant. Experimental results demonstrate that adding discrete representation on four architecture variants strengthens ViT robustness by up to 12% across seven ImageNet robustness benchmarks while maintaining the performance on ImageNet.

We introduce a novel method to teach a robotic agent to interactively explore cluttered yet structured scenes, such as kitchen pantries and grocery shelves, by leveraging the physical plausibility of the scene. We propose a novel learning framework to train an effective scene exploration policy to discover hidden objects with minimal interactions. First, we define a novel scene grammar to represent structured clutter. Then we train a Graph Neural Network (GNN) based Scene Generation agent using deep reinforcement learning (deep RL), to manipulate this Scene Grammar to create a diverse set of stable scenes, each containing multiple hidden objects. Given such cluttered scenes, we then train a Scene Exploration agent, using deep RL, to uncover hidden objects by interactively rearranging the scene. We show that our learned agents hide and discover significantly more objects than the baselines. We present quantitative results that prove the generalization capabilities of our agents. We also demonstrate sim-to-real transfer by successfully deploying the learned policy on a real UR10 robot to explore real-world cluttered scenes. The supplemental video can be found at https://www.youtube.com/watch?v=T2Jo7wwaXss.

In this paper we address the problem of automatically discovering atomic actions in unsupervised manner from instructional videos. Instructional videos contain complex activities and are a rich source of information for intelligent agents, such as, autonomous robots or virtual assistants, which can, for example, automatically `read' the steps from an instructional video and execute them. However, videos are rarely annotated with atomic activities, their boundaries or duration. We present an unsupervised approach to learn atomic actions of structured human tasks from a variety of instructional videos. We propose a sequential stochastic autoregressive model for temporal segmentation of videos, which learns to represent and discover the sequential relationship between different atomic actions of the task, and which provides automatic and unsupervised self-labeling for videos. Our approach outperforms the state-of-the-art unsupervised methods with large margins. We will open source the code., Comment: Full paper

In this paper we address the problem of automatically discovering atomic actions in unsupervised manner from instructional videos, which are rarely annotated with atomic actions. We present an unsupervised approach to learn atomic actions of structured human tasks from a variety of instructional videos based on a sequential stochastic autoregressive model for temporal segmentation of videos. This learns to represent and discover the sequential relationship between different atomic actions of the task, and which provides automatic and unsupervised self-labeling., Comment: 4 page abstract for LUV workshop

We propose an automatic approach that extracts editing styles in a source video and applies the edits to matched footage for video creation. Our Computer Vision based techniques considers framing, content type, playback speed, and lighting of each input video segment. By applying a combination of these features, we demonstrate an effective method that automatically transfers the visual and temporal styles from professionally edited videos to unseen raw footage. We evaluated our approach with real-world videos that contained a total of 3872 video shots of a variety of editing styles, including different subjects, camera motions, and lighting. We reported feedback from survey participants who reviewed a set of our results., Comment: Published to AI for Content Creation Workshop at CVPR 2021

Action prediction focuses on anticipating actions before they happen. Recent works leverage probabilistic approaches to describe future uncertainties and sample future actions. However, these methods cannot easily find all alternative predictions, which are essential given the inherent unpredictability of the future, and current evaluation protocols do not measure a system's ability to find such alternatives. We re-examine action prediction in terms of its ability to predict not only the top predictions, but also top alternatives with the accuracy@k metric. In addition, we propose Choice F1: a metric inspired by F1 score which evaluates a prediction system's ability to find all plausible futures while keeping only the most probable ones. To evaluate this problem, we present a novel method, Predicting the Likely Alternative Next Best, or PLAN-B, for action prediction which automatically finds the set of most likely alternative futures. PLAN-B consists of two novel components: (i) a Choice Table which ensures that all possible futures are found, and (ii) a "Collaborative" RNN system which combines both action sequence and feature information. We demonstrate that our system outperforms state-of-the-art results on benchmark datasets.

PointGoal navigation has seen significant recent interest and progress, spurred on by the Habitat platform and associated challenge. In this paper, we study PointGoal navigation under both a sample budget (75 million frames) and a compute budget (1 GPU for 1 day). We conduct an extensive set of experiments, cumulatively totaling over 50,000 GPU-hours, that let us identify and discuss a number of ostensibly minor but significant design choices -- the advantage estimation procedure (a key component in training), visual encoder architecture, and a seemingly minor hyper-parameter change. Overall, these design choices to lead considerable and consistent improvements over the baselines present in Savva et al. Under a sample budget, performance for RGB-D agents improves 8 SPL on Gibson (14% relative improvement) and 20 SPL on Matterport3D (38% relative improvement). Under a compute budget, performance for RGB-D agents improves by 19 SPL on Gibson (32% relative improvement) and 35 SPL on Matterport3D (220% relative improvement). We hope our findings and recommendations will make serve to make the community's experiments more efficient.