Search

Cache-based approaches stand out as both effective and efficient for adapting vision-language models (VLMs). Nonetheless, the existing cache model overlooks three crucial aspects. 1) Pre-trained VLMs are mainly optimized for image-text similarity, neglecting the importance of image-image similarity, leading to a gap between pre-training and adaptation. 2) The current cache model is based on the Nadaraya-Watson (N-W) estimator, which disregards the intricate relationships among training samples while constructing weight function. 3) Under the condition of limited samples, the logits generated by cache model are of high uncertainty, directly using these logits without accounting for the confidence could be problematic. This work presents three calibration modules aimed at addressing the above challenges. Similarity Calibration refines the image-image similarity by using unlabeled images. We add a learnable projection layer with residual connection on top of the pre-trained image encoder of CLIP and optimize the parameters by minimizing self-supervised contrastive loss. Weight Calibration introduces a precision matrix into the weight function to adequately model the relation between training samples, transforming the existing cache model to a Gaussian Process (GP) regressor, which could be more accurate than N-W estimator. Confidence Calibration leverages the predictive variances computed by GP Regression to dynamically re-scale the logits of cache model, ensuring that the cache model's outputs are appropriately adjusted based on their confidence levels. Besides, to reduce the high complexity of GPs, we further propose a group-based learning strategy. Integrating the above designs, we propose both training-free and training-required variants. Extensive experiments on 11 few-shot classification datasets validate that the proposed methods can achieve state-of-the-art performance., Comment: submitted to IJCV

We propose a generalized method for boosting the generalization ability of pre-trained vision-language models (VLMs) while fine-tuning on downstream few-shot tasks. The idea is realized by exploiting out-of-distribution (OOD) detection to predict whether a sample belongs to a base distribution or a novel distribution and then using the score generated by a dedicated competition based scoring function to fuse the zero-shot and few-shot classifier. The fused classifier is dynamic, which will bias towards the zero-shot classifier if a sample is more likely from the distribution pre-trained on, leading to improved base-to-novel generalization ability. Our method is performed only in test stage, which is applicable to boost existing methods without time-consuming re-training. Extensive experiments show that even weak distribution detectors can still improve VLMs' generalization ability. Specifically, with the help of OOD detectors, the harmonic mean of CoOp and ProGrad increase by 2.6 and 1.5 percentage points over 11 recognition datasets in the base-to-novel setting., Comment: Accepted by AAAI2024

Context Optimization (CoOp) has emerged as a simple yet effective technique for adapting CLIP-like vision-language models to downstream image recognition tasks. Nevertheless, learning compact context with satisfactory base-to-new, domain and cross-task generalization ability while adapting to new tasks is still a challenge. To tackle such a challenge, we propose a lightweight yet generalizable approach termed Compositional Kronecker Context Optimization (CK-CoOp). Technically, the prompt's context words in CK-CoOp are learnable vectors, which are crafted by linearly combining base vectors sourced from a dictionary. These base vectors consist of a non-learnable component obtained by quantizing the weights in the token embedding layer, and a learnable component constructed by applying Kronecker product on several learnable tiny matrices. Intuitively, the compositional structure mitigates the risk of overfitting on training data by remembering more pre-trained knowledge. Meantime, the Kronecker product breaks the non-learnable restrictions of the dictionary, thereby enhancing representation ability with minimal additional parameters. Extensive experiments confirm that CK-CoOp achieves state-of-the-art performance under base-to-new, domain and cross-task generalization evaluation, but also has the metrics of fewer learnable parameters and efficient training and inference speed.

This paper presents Incremental Vision-Language Object Detection (IVLOD), a novel learning task designed to incrementally adapt pre-trained Vision-Language Object Detection Models (VLODMs) to various specialized domains, while simultaneously preserving their zero-shot generalization capabilities for the generalized domain. To address this new challenge, we present the Zero-interference Reparameterizable Adaptation (ZiRa), a novel method that introduces Zero-interference Loss and reparameterization techniques to tackle IVLOD without incurring additional inference costs or a significant increase in memory usage. Comprehensive experiments on COCO and ODinW-13 datasets demonstrate that ZiRa effectively safeguards the zero-shot generalization ability of VLODMs while continuously adapting to new tasks. Specifically, after training on ODinW-13 datasets, ZiRa exhibits superior performance compared to CL-DETR and iDETR, boosting zero-shot generalizability by substantial 13.91 and 8.74 AP, respectively.Our code is available at https://github.com/JarintotionDin/ZiRaGroundingDINO., Comment: This paper has been accepted by NeurIPS 2024

Vision-language models have recently shown great potential on many tasks in computer vision. Meanwhile, prior work demonstrates prompt tuning designed for vision-language models could acquire superior performance on few-shot image recognition compared to linear probe, a strong baseline. In practice, many few-shot tasks are inherently correlated, particularly within specialized domains. However, such information is overlooked previously. Inspired by the fact that modeling task relationship by multi-task learning can usually boost performance, we propose a novel method SoftCPT (Soft Context Sharing for Prompt Tuning) to tune pre-trained vision-language models on multiple target few-shot tasks jointly. Specifically, we design a task-shared meta network to generate prompt context for each task using task name together with a learnable task context as input. The parameters of this meta network as well as the task context are tuned on the joint training set of all tasks. As such, the prompt context of all tasks will be shared in a soft manner. Extensive experiments across four multi-task few-shot datasets covering 44 tasks and 1593 categories demonstrate that SoftCPT significantly outperforms single-task prompt tuning methods, highlighting the effectiveness of multi-task learning for vision-language prompt tuning. Code is available at https://github.com/kding1225/softcpt., Comment: 20 pages

Class incremental learning(CIL) has attracted much attention, but most existing related works focus on fine-tuning the entire representation model, which inevitably results in much catastrophic forgetting. In the contrast, with a semantic-rich pre-trained representation model, parameter-additional-tuning (PAT) only changes very few parameters to learn new visual concepts. Recent studies have proved that PAT-based CIL can naturally avoid fighting against forgetting by replaying or distilling like most of the existing methods. However, we find that PAT-based CIL still faces serious semantic drift, the high-level forgetting problem caused by classifier learning bias at different learning phases, which significantly reduces the performance of PAT-based CIL. To address this problem, we propose Incremental Prototype Tuning (IPT), a simple but effective method that tunes category prototypes for classification and learning example prototypes to compensate for semantic drift. Extensive experiments demonstrate that our method can effectively compensate for semantic drift. Combined with well-pre-trained Vit backbones and other PAT methods, IPT surpasses the state-of-the-art baselines on mainstream incremental learning benchmarks.

Unsupervised domain adaptation (UDA) and semi-supervised learning (SSL) are two typical strategies to reduce expensive manual annotations in machine learning. In order to learn effective models for a target task, UDA utilizes the available labeled source data, which may have different distributions from unlabeled samples in the target domain, while SSL employs few manually annotated target samples. Although UDA and SSL are seemingly very different strategies, we find that they are closely related in terms of task objectives and solutions, and SSL is a special case of UDA problems. Based on this finding, we further investigate whether SSL methods work on UDA tasks. By adapting eight representative SSL algorithms on UDA benchmarks, we show that SSL methods are strong UDA learners. Especially, state-of-the-art SSL methods significantly outperform existing UDA methods on the challenging UDA benchmark of DomainNet, and state-of-the-art UDA methods could be further enhanced with SSL techniques. We thus promote that SSL methods should be employed as baselines in future UDA studies and expect that the revealed relationship between UDA and SSL could shed light on future UDA development. Codes are available at \url{https://github.com/YBZh}., Comment: Codes are available at https://github.com/YBZh

Unsupervised domain adaptation (UDA) aims to learn models for a target domain of unlabeled data by transferring knowledge from a labeled source domain. In the traditional UDA setting, labeled source data are assumed to be available for adaptation. Due to increasing concerns for data privacy, source-free UDA is highly appreciated as a new UDA setting, where only a trained source model is assumed to be available, while labeled source data remain private. However, trained source models may also be unavailable in practice since source models may have commercial values and exposing source models brings risks to the source domain, e.g., problems of model misuse and white-box attacks. In this work, we study a subtly different setting, named Black-Box Unsupervised Domain Adaptation (B$^2$UDA), where only the application programming interface of source model is accessible to the target domain; in other words, the source model itself is kept as a black-box one. To tackle B$^2$UDA, we propose a simple yet effective method, termed Iterative Learning with Noisy Labels (IterLNL). With black-box models as tools of noisy labeling, IterLNL conducts noisy labeling and learning with noisy labels (LNL), iteratively. To facilitate the implementation of LNL in B$^2$UDA, we estimate the noise rate from model predictions of unlabeled target data and propose category-wise sampling to tackle the unbalanced label noise among categories. Experiments on benchmark datasets show the efficacy of IterLNL. Given neither source data nor source models, IterLNL performs comparably with traditional UDA methods that make full use of labeled source data.