Search

Semantic image synthesis, i.e., generating images from user-provided semantic label maps, is an important conditional image generation task as it allows to control both the content as well as the spatial layout of generated images. Although diffusion models have pushed the state of the art in generative image modeling, the iterative nature of their inference process makes them computationally demanding. Other approaches such as GANs are more efficient as they only need a single feed-forward pass for generation, but the image quality tends to suffer on large and diverse datasets. In this work, we propose a new class of GAN discriminators for semantic image synthesis that generates highly realistic images by exploiting feature backbone networks pre-trained for tasks such as image classification. We also introduce a new generator architecture with better context modeling and using cross-attention to inject noise into latent variables, leading to more diverse generated images. Our model, which we dub DP-SIMS, achieves state-of-the-art results in terms of image quality and consistency with the input label maps on ADE-20K, COCO-Stuff, and Cityscapes, surpassing recent diffusion models while requiring two orders of magnitude less compute for inference.

State-of-the-art rehearsal-free continual learning methods exploit the peculiarities of Vision Transformers to learn task-specific prompts, drastically reducing catastrophic forgetting. However, there is a tradeoff between the number of learned parameters and the performance, making such models computationally expensive. In this work, we aim to reduce this cost while maintaining competitive performance. We achieve this by revisiting and extending a simple transfer learning idea: learning task-specific normalization layers. Specifically, we tune the scale and bias parameters of LayerNorm for each continual learning task, selecting them at inference time based on the similarity between task-specific keys and the output of the pre-trained model. To make the classifier robust to incorrect selection of parameters during inference, we introduce a two-stage training procedure, where we first optimize the task-specific parameters and then train the classifier with the same selection procedure of the inference time. Experiments on ImageNet-R and CIFAR-100 show that our method achieves results that are either superior or on par with {the state of the art} while being computationally cheaper., Comment: In The First Workshop on Visual Continual Learning (ICCVW 2023); Oral

Although instance segmentation methods have improved considerably, the dominant paradigm is to rely on fully-annotated training images, which are tedious to obtain. To alleviate this reliance, and boost results, semi-supervised approaches leverage unlabeled data as an additional training signal that limits overfitting to the labeled samples. In this context, we present novel design choices to significantly improve teacher-student distillation models. In particular, we (i) improve the distillation approach by introducing a novel "guided burn-in" stage, and (ii) evaluate different instance segmentation architectures, as well as backbone networks and pre-training strategies. Contrary to previous work which uses only supervised data for the burn-in period of the student model, we also use guidance of the teacher model to exploit unlabeled data in the burn-in period. Our improved distillation approach leads to substantial improvements over previous state-of-the-art results. For example, on the Cityscapes dataset we improve mask-AP from 23.7 to 33.9 when using labels for 10\% of images, and on the COCO dataset we improve mask-AP from 18.3 to 34.1 when using labels for only 1\% of the training data., Comment: Accepted at the 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2024)

Deep convolutional networks are ubiquitous in computer vision, due to their excellent performance across different tasks for various domains. Models are, however, often trained in isolation for each task, failing to exploit relatedness between tasks and domains to learn more compact models that generalise better in low-data regimes. Multi-domain learning aims to handle related tasks, such as image classification across multiple domains, simultaneously. Previous work on this problem explored the use of a pre-trained and fixed domain-agnostic base network, in combination with smaller learnable domain-specific adaptation modules. In this paper, we introduce Modulation Adapters, which update the convolutional filter weights of the model in a multiplicative manner for each task. Parameterising these adaptation weights in a factored manner allows us to scale the number of per-task parameters in a flexible manner, and to strike different parameter-accuracy trade-offs. We evaluate our approach on the Visual Decathlon challenge, composed of ten image classification tasks across different domains, and on the ImageNet-to-Sketch benchmark, which consists of six image classification tasks. Our approach yields excellent results, with accuracies that are comparable to or better than those of existing state-of-the-art approaches.

Self-supervised learning models have been shown to learn rich visual representations without requiring human annotations. However, in many real-world scenarios, labels are partially available, motivating a recent line of work on semi-supervised methods inspired by self-supervised principles. In this paper, we propose a conceptually simple yet empirically powerful approach to turn clustering-based self-supervised methods such as SwAV or DINO into semi-supervised learners. More precisely, we introduce a multi-task framework merging a supervised objective using ground-truth labels and a self-supervised objective relying on clustering assignments with a single cross-entropy loss. This approach may be interpreted as imposing the cluster centroids to be class prototypes. Despite its simplicity, we provide empirical evidence that our approach is highly effective and achieves state-of-the-art performance on CIFAR100 and ImageNet., Comment: CVPR 2023 - Code available at https://github.com/pietroastolfi/suave-daino

The success of transformer models trained with a language modeling objective brings a promising opportunity to the reinforcement learning framework. Decision Transformer is a step towards this direction, showing how to train transformers with a similar next-step prediction objective on offline data. Another important development in this area is the recent emergence of large-scale datasets collected from the internet, such as the ones composed of tutorial videos with captions where people talk about what they are doing. To take advantage of this language component, we propose a novel method for unifying language reasoning with actions in a single policy. Specifically, we augment a transformer policy with word outputs, so it can generate textual captions interleaved with actions. When tested on the most challenging task in BabyAI, with captions describing next subgoals, our reasoning policy consistently outperforms the caption-free baseline.

Developing agents that can execute multiple skills by learning from pre-collected datasets is an important problem in robotics, where online interaction with the environment is extremely time-consuming. Moreover, manually designing reward functions for every single desired skill is prohibitive. Prior works targeted these challenges by learning goal-conditioned policies from offline datasets without manually specified rewards, through hindsight relabelling. These methods suffer from the issue of sparsity of rewards, and fail at long-horizon tasks. In this work, we propose a novel self-supervised learning phase on the pre-collected dataset to understand the structure and the dynamics of the model, and shape a dense reward function for learning policies offline. We evaluate our method on three continuous control tasks, and show that our model significantly outperforms existing approaches, especially on tasks that involve long-term planning., Comment: Code: https://github.com/facebookresearch/go-fresh

Recent image generation models such as Stable Diffusion have exhibited an impressive ability to generate fairly realistic images starting from a simple text prompt. Could such models render real images obsolete for training image prediction models? In this paper, we answer part of this provocative question by investigating the need for real images when training models for ImageNet classification. Provided only with the class names that have been used to build the dataset, we explore the ability of Stable Diffusion to generate synthetic clones of ImageNet and measure how useful these are for training classification models from scratch. We show that with minimal and class-agnostic prompt engineering, ImageNet clones are able to close a large part of the gap between models produced by synthetic images and models trained with real images, for the several standard classification benchmarks that we consider in this study. More importantly, we show that models trained on synthetic images exhibit strong generalization properties and perform on par with models trained on real data for transfer. Project page: https://europe.naverlabs.com/imagenet-sd/, Comment: Accepted to CVPR 2023

Despite significant advances, the performance of state-of-the-art continual learning approaches hinges on the unrealistic scenario of fully labeled data. In this paper, we tackle this challenge and propose an approach for continual semi-supervised learning--a setting where not all the data samples are labeled. A primary issue in this scenario is the model forgetting representations of unlabeled data and overfitting the labeled samples. We leverage the power of nearest-neighbor classifiers to nonlinearly partition the feature space and flexibly model the underlying data distribution thanks to its non-parametric nature. This enables the model to learn a strong representation for the current task, and distill relevant information from previous tasks. We perform a thorough experimental evaluation and show that our method outperforms all the existing approaches by large margins, setting a solid state of the art on the continual semi-supervised learning paradigm. For example, on CIFAR-100 we surpass several others even when using at least 30 times less supervision (0.8% vs. 25% of annotations). Finally, our method works well on both low and high resolution images and scales seamlessly to more complex datasets such as ImageNet-100. The code is publicly available on https://github.com/kangzhiq/NNCSL, Comment: Accepted at ICCV 2023

We propose a novel method to increase shift invariance and prediction accuracy in convolutional neural networks. Specifically, we replace the first-layer combination "real-valued convolutions + max pooling" (RMax) by "complex-valued convolutions + modulus" (CMod), which is stable to translations, or shifts. To justify our approach, we claim that CMod and RMax produce comparable outputs when the convolution kernel is band-pass and oriented (Gabor-like filter). In this context, CMod can therefore be considered as a stable alternative to RMax. To enforce this property, we constrain the convolution kernels to adopt such a Gabor-like structure. The corresponding architecture is called mathematical twin, because it employs a well-defined mathematical operator to mimic the behavior of the original, freely-trained model. Our approach achieves superior accuracy on ImageNet and CIFAR-10 classification tasks, compared to prior methods based on low-pass filtering. Arguably, our approach's emphasis on retaining high-frequency details contributes to a better balance between shift invariance and information preservation, resulting in improved performance. Furthermore, it has a lower computational cost and memory footprint than concurrent work, making it a promising solution for practical implementation.

Vision Transformers (ViTs) have become a dominant paradigm for visual representation learning with self-attention operators. Although these operators provide flexibility to the model with their adjustable attention kernels, they suffer from inherent limitations: (1) the attention kernel is not discriminative enough, resulting in high redundancy of the ViT layers, and (2) the complexity in computation and memory is quadratic in the sequence length. In this paper, we propose a novel attention operator, called lightweight structure-aware attention (LiSA), which has a better representation power with log-linear complexity. Our operator learns structural patterns by using a set of relative position embeddings (RPEs). To achieve log-linear complexity, the RPEs are approximated with fast Fourier transforms. Our experiments and ablation studies demonstrate that ViTs based on the proposed operator outperform self-attention and other existing operators, achieving state-of-the-art results on ImageNet, and competitive results on other visual understanding benchmarks such as COCO and Something-Something-V2. The source code of our approach will be released online., Comment: 8 pages, 5 figures

In defense-related remote sensing applications, such as vehicle detection on satellite imagery, supervised learning requires a huge number of labeled examples to reach operational performances. Such data are challenging to obtain as it requires military experts, and some observables are intrinsically rare. This limited labeling capability, as well as the large number of unlabeled images available due to the growing number of sensors, make object detection on remote sensing imagery highly relevant for self-supervised learning. We study in-domain self-supervised representation learning for object detection on very high resolution optical satellite imagery, that is yet poorly explored. For the first time to our knowledge, we study the problem of label efficiency on this task. We use the large land use classification dataset Functional Map of the World to pretrain representations with an extension of the Momentum Contrast framework. We then investigate this model's transferability on a real-world task of fine-grained vehicle detection and classification on Preligens proprietary data, which is designed to be representative of an operational use case of strategic site surveillance. We show that our in-domain self-supervised learning model is competitive with ImageNet pretraining, and outperforms it in the low-label regime.

The application of deep neural networks to remote sensing imagery is often constrained by the lack of ground-truth annotations. Adressing this issue requires models that generalize efficiently from limited amounts of labeled data, allowing us to tackle a wider range of Earth observation tasks. Another challenge in this domain is developing algorithms that operate at variable spatial resolutions, e.g., for the problem of classifying land use at different scales. Recently, self-supervised learning has been applied in the remote sensing domain to exploit readily-available unlabeled data, and was shown to reduce or even close the gap with supervised learning. In this paper, we study self-supervised visual representation learning through the lens of label efficiency, for the task of land use classification on multi-resolution/multi-scale satellite images. We benchmark two contrastive self-supervised methods adapted from Momentum Contrast (MoCo) and provide evidence that these methods can be perform effectively given little downstream supervision, where randomly initialized networks fail to generalize. Moreover, they outperform out-of-domain pretraining alternatives. We use the large-scale fMoW dataset to pretrain and evaluate the networks, and validate our observations with transfer to the RESISC45 dataset.

This paper focuses on improving the mathematical interpretability of convolutional neural networks (CNNs) in the context of image classification. Specifically, we tackle the instability issue arising in their first layer, which tends to learn parameters that closely resemble oriented band-pass filters when trained on datasets like ImageNet. Subsampled convolutions with such Gabor-like filters are prone to aliasing, causing sensitivity to small input shifts. In this context, we establish conditions under which the max pooling operator approximates a complex modulus, which is nearly shift invariant. We then derive a measure of shift invariance for subsampled convolutions followed by max pooling. In particular, we highlight the crucial role played by the filter's frequency and orientation in achieving stability. We experimentally validate our theory by considering a deterministic feature extractor based on the dual-tree complex wavelet packet transform, a particular case of discrete Gabor-like decomposition.

We consider the problem of training a deep neural network on a given classification task, e.g., ImageNet-1K (IN1K), so that it excels at both the training task as well as at other (future) transfer tasks. These two seemingly contradictory properties impose a trade-off between improving the model's generalization and maintaining its performance on the original task. Models trained with self-supervised learning tend to generalize better than their supervised counterparts for transfer learning; yet, they still lag behind supervised models on IN1K. In this paper, we propose a supervised learning setup that leverages the best of both worlds. We extensively analyze supervised training using multi-scale crops for data augmentation and an expendable projector head, and reveal that the design of the projector allows us to control the trade-off between performance on the training task and transferability. We further replace the last layer of class weights with class prototypes computed on the fly using a memory bank and derive two models: t-ReX that achieves a new state of the art for transfer learning and outperforms top methods such as DINO and PAWS on IN1K, and t-ReX* that matches the highly optimized RSB-A1 model on IN1K while performing better on transfer tasks. Code and pretrained models: https://europe.naverlabs.com/t-rex, Comment: Accepted to ICLR 2023 (spotlight)

Recent works in autonomous driving have widely adopted the bird's-eye-view (BEV) semantic map as an intermediate representation of the world. Online prediction of these BEV maps involves non-trivial operations such as multi-camera data extraction as well as fusion and projection into a common topview grid. This is usually done with error-prone geometric operations (e.g., homography or back-projection from monocular depth estimation) or expensive direct dense mapping between image pixels and pixels in BEV (e.g., with MLP or attention). In this work, we present 'LaRa', an efficient encoder-decoder, transformer-based model for vehicle semantic segmentation from multiple cameras. Our approach uses a system of cross-attention to aggregate information over multiple sensors into a compact, yet rich, collection of latent representations. These latent representations, after being processed by a series of self-attention blocks, are then reprojected with a second cross-attention in the BEV space. We demonstrate that our model outperforms the best previous works using transformers on nuScenes. The code and trained models are available at https://github.com/valeoai/LaRa

Learning a diverse set of skills by interacting with an environment without any external supervision is an important challenge. In particular, obtaining a goal-conditioned agent that can reach any given state is useful in many applications. We propose a novel method for training such a goal-conditioned agent without any external rewards or any domain knowledge. We use random walk to train a reachability network that predicts the similarity between two states. This reachability network is then used in building goal memory containing past observations that are diverse and well-balanced. Finally, we train a goal-conditioned policy network with goals sampled from the goal memory and reward it by the reachability network and the goal memory. All the components are kept updated throughout training as the agent discovers and learns new goals. We apply our method to a continuous control navigation and robotic manipulation tasks.

Audio-visual automatic speech recognition (AV-ASR) is an extension of ASR that incorporates visual cues, often from the movements of a speaker's mouth. Unlike works that simply focus on the lip motion, we investigate the contribution of entire visual frames (visual actions, objects, background etc.). This is particularly useful for unconstrained videos, where the speaker is not necessarily visible. To solve this task, we propose a new sequence-to-sequence AudioVisual ASR TrAnsformeR (AVATAR) which is trained end-to-end from spectrograms and full-frame RGB. To prevent the audio stream from dominating training, we propose different word-masking strategies, thereby encouraging our model to pay attention to the visual stream. We demonstrate the contribution of the visual modality on the How2 AV-ASR benchmark, especially in the presence of simulated noise, and show that our model outperforms all other prior work by a large margin. Finally, we also create a new, real-world test bed for AV-ASR called VisSpeech, which demonstrates the contribution of the visual modality under challenging audio conditions.

Both a good understanding of geometrical concepts and a broad familiarity with objects lead to our excellent perception of moving objects. The human ability to detect and segment moving objects works in the presence of multiple objects, complex background geometry, motion of the observer and even camouflage. How humans perceive moving objects so reliably is a longstanding research question in computer vision and borrows findings from related areas such as psychology, cognitive science and physics. One approach to the problem is to teach a deep network to model all of these effects. This contrasts with the strategy used by human vision, where cognitive processes and body design are tightly coupled and each is responsible for certain aspects of correctly identifying moving objects. Similarly from the computer vision perspective, there is evidence that classical, geometry-based techniques are better suited to the "motion-based" parts of the problem, while deep networks are more suitable for modeling appearance. In this work, we argue that the coupling of camera rotation and camera translation can create complex motion fields that are difficult for a deep network to untangle directly. We present a novel probabilistic model to estimate the camera's rotation given the motion field. We then rectify the flow field to obtain a rotation-compensated motion field for subsequent segmentation. This strategy of first estimating camera motion, and then allowing a network to learn the remaining parts of the problem, yields improved results on the widely used DAVIS benchmark as well as the recently published motion segmentation data set MoCA (Moving Camouflaged Animals).

Self-supervised models have been shown to produce comparable or better visual representations than their supervised counterparts when trained offline on unlabeled data at scale. However, their efficacy is catastrophically reduced in a Continual Learning (CL) scenario where data is presented to the model sequentially. In this paper, we show that self-supervised loss functions can be seamlessly converted into distillation mechanisms for CL by adding a predictor network that maps the current state of the representations to their past state. This enables us to devise a framework for Continual self-supervised visual representation Learning that (i) significantly improves the quality of the learned representations, (ii) is compatible with several state-of-the-art self-supervised objectives, and (iii) needs little to no hyperparameter tuning. We demonstrate the effectiveness of our approach empirically by training six popular self-supervised models in various CL settings.

Pre-training on large scale unlabelled datasets has shown impressive performance improvements in the fields of computer vision and natural language processing. Given the advent of large-scale instructional video datasets, a common strategy for pre-training video encoders is to use the accompanying speech as weak supervision. However, as speech is used to supervise the pre-training, it is never seen by the video encoder, which does not learn to process that modality. We address this drawback of current pre-training methods, which fail to exploit the rich cues in spoken language. Our proposal is to pre-train a video encoder using all the available video modalities as supervision, namely, appearance, sound, and transcribed speech. We mask an entire modality in the input and predict it using the other two modalities. This encourages each modality to collaborate with the others, and our video encoder learns to process appearance and audio as well as speech. We show the superior performance of our "modality masking" pre-training approach for video retrieval on the How2R, YouCook2 and Condensed Movies datasets., Comment: Accepted at WACV 2022

We introduce regularized Frank-Wolfe, a general and effective algorithm for inference and learning of dense conditional random fields (CRFs). The algorithm optimizes a nonconvex continuous relaxation of the CRF inference problem using vanilla Frank-Wolfe with approximate updates, which are equivalent to minimizing a regularized energy function. Our proposed method is a generalization of existing algorithms such as mean field or concave-convex procedure. This perspective not only offers a unified analysis of these algorithms, but also allows an easy way of exploring different variants that potentially yield better performance. We illustrate this in our empirical results on standard semantic segmentation datasets, where several instantiations of our regularized Frank-Wolfe outperform mean field inference, both as a standalone component and as an end-to-end trainable layer in a neural network. We also show that dense CRFs, coupled with our new algorithms, produce significant improvements over strong CNN baselines., Comment: NeurIPS 2021. This version fixed some minor typos (constant factor 2 removed from bottom-right cell of Theorem 1's table, and from last row of Table 5)

Vision-based depth estimation is a key feature in autonomous systems, which often relies on a single camera or several independent ones. In such a monocular setup, dense depth is obtained with either additional input from one or several expensive LiDARs, e.g., with 64 beams, or camera-only methods, which suffer from scale-ambiguity and infinite-depth problems. In this paper, we propose a new alternative of densely estimating metric depth by combining a monocular camera with a light-weight LiDAR, e.g., with 4 beams, typical of today's automotive-grade mass-produced laser scanners. Inspired by recent self-supervised methods, we introduce a novel framework, called LiDARTouch, to estimate dense depth maps from monocular images with the help of ``touches'' of LiDAR, i.e., without the need for dense ground-truth depth. In our setup, the minimal LiDAR input contributes on three different levels: as an additional model's input, in a self-supervised LiDAR reconstruction objective function, and to estimate changes of pose (a key component of self-supervised depth estimation architectures). Our LiDARTouch framework achieves new state of the art in self-supervised depth estimation on the KITTI dataset, thus supporting our choices of integrating the very sparse LiDAR signal with other visual features. Moreover, we show that the use of a few-beam LiDAR alleviates scale ambiguity and infinite-depth issues that camera-only methods suffer from. We also demonstrate that methods from the fully-supervised depth-completion literature can be adapted to a self-supervised regime with a minimal LiDAR signal.

In this work, we present a memory-augmented approach for image-goal navigation. Earlier attempts, including RL-based and SLAM-based approaches have either shown poor generalization performance, or are heavily-reliant on pose/depth sensors. Our method is based on an attention-based end-to-end model that leverages an episodic memory to learn to navigate. First, we train a state-embedding network in a self-supervised fashion, and then use it to embed previously-visited states into the agent's memory. Our navigation policy takes advantage of this information through an attention mechanism. We validate our approach with extensive evaluations, and show that our model establishes a new state of the art on the challenging Gibson dataset. Furthermore, we achieve this impressive performance from RGB input alone, without access to additional information such as position or depth, in stark contrast to related work.

Measuring concept generalization, i.e., the extent to which models trained on a set of (seen) visual concepts can be leveraged to recognize a new set of (unseen) concepts, is a popular way of evaluating visual representations, especially in a self-supervised learning framework. Nonetheless, the choice of unseen concepts for such an evaluation is usually made arbitrarily, and independently from the seen concepts used to train representations, thus ignoring any semantic relationships between the two. In this paper, we argue that the semantic relationships between seen and unseen concepts affect generalization performance and propose ImageNet-CoG, a novel benchmark on the ImageNet-21K (IN-21K) dataset that enables measuring concept generalization in a principled way. Our benchmark leverages expert knowledge that comes from WordNet in order to define a sequence of unseen IN-21K concept sets that are semantically more and more distant from the ImageNet-1K (IN-1K) subset, a ubiquitous training set. This allows us to benchmark visual representations learned on IN-1K out-of-the box. We conduct a large-scale study encompassing 31 convolution and transformer-based models and show how different architectures, levels of supervision, regularization techniques and use of web data impact the concept generalization performance., Comment: Accepted to ICCV 2021. See our project website: https://europe.naverlabs.com/cog-benchmark for code and ImageNet-CoG level files

We propose a novel amortized variational inference scheme for an empirical Bayes meta-learning model, where model parameters are treated as latent variables. We learn the prior distribution over model parameters conditioned on limited training data using a variational autoencoder approach. Our framework proposes sharing the same amortized inference network between the conditional prior and variational posterior distributions over the model parameters. While the posterior leverages both the labeled support and query data, the conditional prior is based only on the labeled support data. We show that in earlier work, relying on Monte-Carlo approximation, the conditional prior collapses to a Dirac delta function. In contrast, our variational approach prevents this collapse and preserves uncertainty over the model parameters. We evaluate our approach on the miniImageNet, CIFAR-FS and FC100 datasets, and present results demonstrating its advantages over previous work., Comment: ICML 2020

We present a new video understanding pentathlon challenge, an open competition held in conjunction with the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2020. The objective of the challenge was to explore and evaluate new methods for text-to-video retrieval-the task of searching for content within a corpus of videos using natural language queries. This report summarizes the results of the first edition of the challenge together with the findings of the participants., Comment: Individual reports, dataset information, rules, and released source code can be found at the competition webpage (https://www.robots.ox.ac.uk/~vgg/challenges/video-pentathlon)

The task of retrieving video content relevant to natural language queries plays a critical role in effectively handling internet-scale datasets. Most of the existing methods for this caption-to-video retrieval problem do not fully exploit cross-modal cues present in video. Furthermore, they aggregate per-frame visual features with limited or no temporal information. In this paper, we present a multi-modal transformer to jointly encode the different modalities in video, which allows each of them to attend to the others. The transformer architecture is also leveraged to encode and model the temporal information. On the natural language side, we investigate the best practices to jointly optimize the language embedding together with the multi-modal transformer. This novel framework allows us to establish state-of-the-art results for video retrieval on three datasets. More details are available at http://thoth.inrialpes.fr/research/MMT., Comment: ECCV 2020 (spotlight paper)

Eye movement and strategic placement of the visual field onto the retina, gives animals increased resolution of the scene and suppresses distracting information. This fundamental system has been missing from video understanding with deep networks, typically limited to 224 by 224 pixel content locked to the camera frame. We propose a simple idea, WorldFeatures, where each feature at every layer has a spatial transformation, and the feature map is only transformed as needed. We show that a network built with these WorldFeatures, can be used to model eye movements, such as saccades, fixation, and smooth pursuit, even in a batch setting on pre-recorded video. That is, the network can for example use all 224 by 224 pixels to look at a small detail one moment, and the whole scene the next. We show that typical building blocks, such as convolutions and pooling, can be adapted to support WorldFeatures using available tools. Experiments are presented on the Charades, Olympic Sports, and Caltech-UCSD Birds-200-2011 datasets, exploring action recognition, fine-grained recognition, and video stabilization.

Unsupervised learning of generative models has seen tremendous progress over recent years, in particular due to generative adversarial networks (GANs), variational autoencoders, and flow-based models. GANs have dramatically improved sample quality, but suffer from two drawbacks: (i) they mode-drop, i.e., do not cover the full support of the train data, and (ii) they do not allow for likelihood evaluations on held-out data. In contrast, likelihood-based training encourages models to cover the full support of the train data, but yields poorer samples. These mutual shortcomings can in principle be addressed by training generative latent variable models in a hybrid adversarial-likelihood manner. However, we show that commonly made parametric assumptions create a conflict between them, making successful hybrid models non trivial. As a solution, we propose to use deep invertible transformations in the latent variable decoder. This approach allows for likelihood computations in image space, is more efficient than fully invertible models, and can take full advantage of adversarial training. We show that our model significantly improves over existing hybrid models: offering GAN-like samples, IS and FID scores that are competitive with fully adversarial models, and improved likelihood scores.

Generative adversarial networks (GANs) are one of the most popular methods for generating images today. While impressive results have been validated by visual inspection, a number of quantitative criteria have emerged only recently. We argue here that the existing ones are insufficient and need to be in adequation with the task at hand. In this paper we introduce two measures based on image classification---GAN-train and GAN-test, which approximate the recall (diversity) and precision (quality of the image) of GANs respectively. We evaluate a number of recent GAN approaches based on these two measures and demonstrate a clear difference in performance. Furthermore, we observe that the increasing difficulty of the dataset, from CIFAR10 over CIFAR100 to ImageNet, shows an inverse correlation with the quality of the GANs, as clearly evident from our measures., Comment: Accepted to ECCV2018

Although deep learning approaches have stood out in recent years due to their state-of-the-art results, they continue to suffer from catastrophic forgetting, a dramatic decrease in overall performance when training with new classes added incrementally. This is due to current neural network architectures requiring the entire dataset, consisting of all the samples from the old as well as the new classes, to update the model -a requirement that becomes easily unsustainable as the number of classes grows. We address this issue with our approach to learn deep neural networks incrementally, using new data and only a small exemplar set corresponding to samples from the old classes. This is based on a loss composed of a distillation measure to retain the knowledge acquired from the old classes, and a cross-entropy loss to learn the new classes. Our incremental training is achieved while keeping the entire framework end-to-end, i.e., learning the data representation and the classifier jointly, unlike recent methods with no such guarantees. We evaluate our method extensively on the CIFAR-100 and ImageNet (ILSVRC 2012) image classification datasets, and show state-of-the-art performance., Comment: To appear in ECCV 2018

Several theories in cognitive neuroscience suggest that when people interact with the world, or simulate interactions, they do so from a first-person egocentric perspective, and seamlessly transfer knowledge between third-person (observer) and first-person (actor). Despite this, learning such models for human action recognition has not been achievable due to the lack of data. This paper takes a step in this direction, with the introduction of Charades-Ego, a large-scale dataset of paired first-person and third-person videos, involving 112 people, with 4000 paired videos. This enables learning the link between the two, actor and observer perspectives. Thereby, we address one of the biggest bottlenecks facing egocentric vision research, providing a link from first-person to the abundant third-person data on the web. We use this data to learn a joint representation of first and third-person videos, with only weak supervision, and show its effectiveness for transferring knowledge from the third-person to the first-person domain., Comment: CVPR 2018 spotlight presentation

In Actor and Observer we introduced a dataset linking the first and third-person video understanding domains, the Charades-Ego Dataset. In this paper we describe the egocentric aspect of the dataset and present annotations for Charades-Ego with 68,536 activity instances in 68.8 hours of first and third-person video, making it one of the largest and most diverse egocentric datasets available. Charades-Ego furthermore shares activity classes, scripts, and methodology with the Charades dataset, that consist of additional 82.3 hours of third-person video with 66,500 activity instances. Charades-Ego has temporal annotations and textual descriptions, making it suitable for egocentric video classification, localization, captioning, and new tasks utilizing the cross-modal nature of the data.

We study the problem of segmenting moving objects in unconstrained videos. Given a video, the task is to segment all the objects that exhibit independent motion in at least one frame. We formulate this as a learning problem and design our framework with three cues: (i) independent object motion between a pair of frames, which complements object recognition, (ii) object appearance, which helps to correct errors in motion estimation, and (iii) temporal consistency, which imposes additional constraints on the segmentation. The framework is a two-stream neural network with an explicit memory module. The two streams encode appearance and motion cues in a video sequence respectively, while the memory module captures the evolution of objects over time, exploiting the temporal consistency. The motion stream is a convolutional neural network trained on synthetic videos to segment independently moving objects in the optical flow field. The module to build a 'visual memory' in video, i.e., a joint representation of all the video frames, is realized with a convolutional recurrent unit learned from a small number of training video sequences. For every pixel in a frame of a test video, our approach assigns an object or background label based on the learned spatio-temporal features as well as the 'visual memory' specific to the video. We evaluate our method extensively on three benchmarks, DAVIS, Freiburg-Berkeley motion segmentation dataset and SegTrack. In addition, we provide an extensive ablation study to investigate both the choice of the training data and the influence of each component in the proposed framework., Comment: arXiv admin note: text overlap with arXiv:1704.05737, arXiv:1612.07217

Despite their success for object detection, convolutional neural networks are ill-equipped for incremental learning, i.e., adapting the original model trained on a set of classes to additionally detect objects of new classes, in the absence of the initial training data. They suffer from "catastrophic forgetting" - an abrupt degradation of performance on the original set of classes, when the training objective is adapted to the new classes. We present a method to address this issue, and learn object detectors incrementally, when neither the original training data nor annotations for the original classes in the new training set are available. The core of our proposed solution is a loss function to balance the interplay between predictions on the new classes and a new distillation loss which minimizes the discrepancy between responses for old classes from the original and the updated networks. This incremental learning can be performed multiple times, for a new set of classes in each step, with a moderate drop in performance compared to the baseline network trained on the ensemble of data. We present object detection results on the PASCAL VOC 2007 and COCO datasets, along with a detailed empirical analysis of the approach., Comment: To appear in ICCV 2017

In this paper, we propose a new framework for action localization that tracks people in videos and extracts full-body human tubes, i.e., spatio-temporal regions localizing actions, even in the case of occlusions or truncations. This is achieved by training a novel human part detector that scores visible parts while regressing full-body bounding boxes. The core of our method is a convolutional neural network which learns part proposals specific to certain body parts. These are then combined to detect people robustly in each frame. Our tracking algorithm connects the image detections temporally to extract full-body human tubes. We apply our new tube extraction method on the problem of human action localization, on the popular JHMDB dataset, and a very recent challenging dataset DALY (Daily Action Localization in YouTube), showing state-of-the-art results., Comment: BMVC 2017

This paper addresses the task of segmenting moving objects in unconstrained videos. We introduce a novel two-stream neural network with an explicit memory module to achieve this. The two streams of the network encode spatial and temporal features in a video sequence respectively, while the memory module captures the evolution of objects over time. The module to build a "visual memory" in video, i.e., a joint representation of all the video frames, is realized with a convolutional recurrent unit learned from a small number of training video sequences. Given a video frame as input, our approach assigns each pixel an object or background label based on the learned spatio-temporal features as well as the "visual memory" specific to the video, acquired automatically without any manually-annotated frames. The visual memory is implemented with convolutional gated recurrent units, which allows to propagate spatial information over time. We evaluate our method extensively on two benchmarks, DAVIS and Freiburg-Berkeley motion segmentation datasets, and show state-of-the-art results. For example, our approach outperforms the top method on the DAVIS dataset by nearly 6%. We also provide an extensive ablative analysis to investigate the influence of each component in the proposed framework.

The problem of determining whether an object is in motion, irrespective of camera motion, is far from being solved. We address this challenging task by learning motion patterns in videos. The core of our approach is a fully convolutional network, which is learned entirely from synthetic video sequences, and their ground-truth optical flow and motion segmentation. This encoder-decoder style architecture first learns a coarse representation of the optical flow field features, and then refines it iteratively to produce motion labels at the original high-resolution. We further improve this labeling with an objectness map and a conditional random field, to account for errors in optical flow, and also to focus on moving "things" rather than "stuff". The output label of each pixel denotes whether it has undergone independent motion, i.e., irrespective of camera motion. We demonstrate the benefits of this learning framework on the moving object segmentation task, where the goal is to segment all objects in motion. Our approach outperforms the top method on the recently released DAVIS benchmark dataset, comprising real-world sequences, by 5.6%. We also evaluate on the Berkeley motion segmentation database, achieving state-of-the-art results.

Fully convolutional neural networks (FCNNs) trained on a large number of images with strong pixel-level annotations have become the new state of the art for the semantic segmentation task. While there have been recent attempts to learn FCNNs from image-level weak annotations, they need additional constraints, such as the size of an object, to obtain reasonable performance. To address this issue, we present motion-CNN (M-CNN), a novel FCNN framework which incorporates motion cues and is learned from video-level weak annotations. Our learning scheme to train the network uses motion segments as soft constraints, thereby handling noisy motion information. When trained on weakly-annotated videos, our method outperforms the state-of-the-art EM-Adapt approach on the PASCAL VOC 2012 image segmentation benchmark. We also demonstrate that the performance of M-CNN learned with 150 weak video annotations is on par with state-of-the-art weakly-supervised methods trained with thousands of images. Finally, M-CNN substantially outperforms recent approaches in a related task of video co-localization on the YouTube-Objects dataset., Comment: Extended version of our ECCV 2016 paper

Recognizing scene text is a challenging problem, even more so than the recognition of scanned documents. This problem has gained significant attention from the computer vision community in recent years, and several methods based on energy minimization frameworks and deep learning approaches have been proposed. In this work, we focus on the energy minimization framework and propose a model that exploits both bottom-up and top-down cues for recognizing cropped words extracted from street images. The bottom-up cues are derived from individual character detections from an image. We build a conditional random field model on these detections to jointly model the strength of the detections and the interactions between them. These interactions are top-down cues obtained from a lexicon-based prior, i.e., language statistics. The optimal word represented by the text image is obtained by minimizing the energy function corresponding to the random field model. We evaluate our proposed algorithm extensively on a number of cropped scene text benchmark datasets, namely Street View Text, ICDAR 2003, 2011 and 2013 datasets, and IIIT 5K-word, and show better performance than comparable methods. We perform a rigorous analysis of all the steps in our approach and analyze the results. We also show that state-of-the-art convolutional neural network features can be integrated in our framework to further improve the recognition performance.

Tracking-by-detection approaches are some of the most successful object trackers in recent years. Their success is largely determined by the detector model they learn initially and then update over time. However, under challenging conditions where an object can undergo transformations, e.g., severe rotation, these methods are found to be lacking. In this paper, we address this problem by formulating it as a proposal selection task and making two contributions. The first one is introducing novel proposals estimated from the geometric transformations undergone by the object, and building a rich candidate set for predicting the object location. The second one is devising a novel selection strategy using multiple cues, i.e., detection score and edgeness score computed from state-of-the-art object edges and motion boundaries. We extensively evaluate our approach on the visual object tracking 2014 challenge and online tracking benchmark datasets, and show the best performance., Comment: ICCV 2015

We present alpha-expansion beta-shrink moves, a simple generalization of the widely-used alpha-beta swap and alpha-expansion algorithms for approximate energy minimization. We show that in a certain sense, these moves dominate both alpha-beta-swap and alpha-expansion moves, but unlike previous generalizations the new moves require no additional assumptions and are still solvable in polynomial-time. We show promising experimental results with the new moves, which we believe could be used in any context where alpha-expansions are currently employed., Comment: Conference on Uncertainty in Artificial Intelligence (2011)