CVPR 2024 论文列表

Despite advancements in text-to-image generation (T2I) prior methods often face text-image misalignment problems such as relation confusion in generated images. Existing solutions involve cross-attention manipulation for better compositional understanding or integrating large language models for improved layout planning. However the inherent alignment capabilities of T2I models are still inadequate. By reviewing the link between generative and discriminative modeling we posit that T2I models' discriminative abilities may reflect their text-image alignment proficiency during generation. In this light we advocate bolstering the discriminative abilities of T2I models to achieve more precise text-to-image alignment for generation. We present a discriminative adapter built on T2I models to probe their discriminative abilities on two representative tasks and leverage discriminative fine-tuning to improve their text-image alignment. As a bonus of the discriminative adapter a self-correction mechanism can leverage discriminative gradients to better align generated images to text prompts during inference. Comprehensive evaluations across three benchmark datasets including both in-distribution and out-of-distribution scenarios demonstrate our method's superior generation performance. Meanwhile it achieves state-of-the-art discriminative performance on the two discriminative tasks compared to other generative models. The code is available at https://dpt-t2i.github.io/.

Addressing biases in computer vision models is crucial for real-world AI deployments. However mitigating visual biases is challenging due to their unexplainable nature often identified indirectly through visualization or sample statistics which necessitates additional human supervision for interpretation. To tackle this issue we propose the Bias-to-Text (B2T) framework which interprets visual biases as keywords. Specifically we extract common keywords from the captions of mispredicted images to identify potential biases in the model. We then validate these keywords by measuring their similarity to the mispredicted images using a vision-language scoring model. The keyword explanation form of visual bias offers several advantages such as a clear group naming for bias discovery and a natural extension for debiasing using these group names. Our experiments demonstrate that B2T can identify known biases such as gender bias in CelebA background bias in Waterbirds and distribution shifts in ImageNet-R/C. Additionally B2T uncovers novel biases in larger datasets such as Dollar Street and ImageNet. For example we discovered a contextual bias between \keyword bee and \keyword flower in ImageNet. We also highlight various applications of B2T keywords including debiased training CLIP prompting and model comparison.

Recently Vision-Language Model (VLM) has greatly advanced the Human-Object Interaction (HOI) detection. The existing VLM-based HOI detectors typically adopt a hand-crafted template (e.g. a photo of a person [action] a/an [object]) to acquire text knowledge through the VLM text encoder. However such approaches only encoding the action-specific text prompts in vocabulary level may suffer from learning ambiguity without exploring the fine-grained clues from the perspective of interaction context. In this paper we propose a novel method to discover Syntactic Interaction Clues for HOI detection (SICHOI) by using VLM. Specifically we first investigate what are the essential elements for an interaction context and then establish a syntactic interaction bank from three levels: spatial relationship action-oriented posture and situational condition. Further to align visual features with the syntactic interaction bank we adopt a multi-view extractor to jointly aggregate visual features from instance interaction and image levels accordingly. In addition we also introduce a dual cross-attention decoder to perform context propagation between text knowledge and visual features thereby enhancing the HOI detection. Experimental results demonstrate that our proposed method achieves state-of-the-art performance on HICO-DET and V-COCO.

Machine learning models can perform well on in-distribution data but often fail on biased subgroups that are underrepresented in the training data hindering the robustness of models for reliable applications. Such subgroups are typically unknown due to the absence of subgroup labels. Discovering biased subgroups is the key to understanding models' failure modes and further improving models' robustness. Most previous works of subgroup discovery make an implicit assumption that models only underperform on a single biased subgroup which does not hold on in-the-wild data where multiple biased subgroups exist. In this work we propose Decomposition Interpretation and Mitigation (DIM) a novel method to address a more challenging but also more practical problem of discovering multiple biased subgroups in image classifiers. Our approach decomposes the image features into multiple components that represent multiple subgroups. This decomposition is achieved via a bilinear dimension reduction method Partial Least Square (PLS) guided by useful supervision from the image classifier. We further interpret the semantic meaning of each subgroup component by generating natural language descriptions using vision-language foundation models. Finally DIM mitigates multiple biased subgroups simultaneously via two strategies including the data- and model-centric strategies. Extensive experiments on CIFAR-100 and Breeds datasets demonstrate the effectiveness of DIM in discovering and mitigating multiple biased subgroups. Furthermore DIM uncovers the failure modes of the classifier on Hard ImageNet showcasing its broader applicability to understanding model bias in image classifiers.

Multimodal learning has advanced the performance for many vision-language tasks. However most existing works in embodied dialog research focus on navigation and leave the localization task understudied. The few existing dialog-based localization approaches assume the availability of entire dialog prior to localizaiton which is impractical for deployed dialog-based localization. In this paper we propose DiaLoc a new dialog-based localization framework which aligns with a real human operator behavior. Specifically we produce an iterative refinement of location predictions which can visualize current pose believes after each dialog turn. DiaLoc effectively utilizes the multimodal data for multi-shot localization where a fusion encoder fuses vision and dialog information iteratively. We achieve state-of-the-art results on embodied dialog-based localization task in single-shot (+7.08% in Acc5@valUnseen) and multi-shot settings (+10.85% in Acc5@valUnseen). DiaLoc narrows the gap between simulation and real-world applications opening doors for future research on collaborative localization and navigation.

We introduce the video detours problem for navigating instructional videos. Given a source video and a natural language query asking to alter the how-to video's current path of execution in a certain way the goal is to find a related "detour video" that satisfies the requested alteration. To address this challenge we propose VidDetours a novel video-language approach that learns to retrieve the targeted temporal segments from a large repository of how-to's using video-and-text conditioned queries. Furthermore we devise a language-based pipeline that exploits how-to video narration text to create weakly supervised training data. We demonstrate our idea applied to the domain of how-to cooking videos where a user can detour from their current recipe to find steps with alternate ingredients tools and techniques. Validating on a ground truth annotated dataset of 16K samples we show our model's significant improvements over best available methods for video retrieval and question answering with recall rates exceeding the state of the art by 35%.

Over the past year a large body of multimodal research has emerged around zero-shot evaluation using GPT descriptors. These studies boost the zero-shot accuracy of pretrained VL models with an ensemble of label-specific text generated by GPT. A recent study WaffleCLIP demonstrated that similar zero-shot accuracy can be achieved with an ensemble of random descriptors. However both zero-shot methods are un-trainable and consequently sub-optimal when some few-shot out-of-distribution (OOD) training data is available. Inspired by these prior works we present two more flexible methods called descriptor and word soups which do not require an LLM at test time and can leverage training data to increase OOD target accuracy. Descriptor soup greedily selects a small set of textual descriptors using generic few-shot training data then calculates robust class embeddings using the selected descriptors. Word soup greedily assembles a chain of words in a similar manner. Compared to existing few-shot soft prompt tuning methods word soup requires fewer parameters by construction and less GPU memory since it does not require backpropagation. Both soups outperform current published few-shot methods even when combined with SoTA zero-shot methods on cross-dataset and domain generalization benchmarks. Compared with SoTA prompt and descriptor ensembling methods such as ProDA and WaffleCLIP word soup achieves higher OOD accuracy with fewer ensemble members. Please checkout our code: https://github.com/Chris210634/word_soups

How do two sets of images differ? Discerning set-level differences is crucial for understanding model behaviors and analyzing datasets yet manually sifting through thousands of images is impractical. To aid in this discovery process we explore the task of automatically describing the differences between two sets of images which we term Set Difference Captioning. This task takes in image sets \mathcal D _A and \mathcal D _B and outputs a description that is more often true on \mathcal D _A than \mathcal D _B. We outline a two-stage approach that first proposes candidate difference descriptions from image sets and then re-ranks the candidates by checking how well they can differentiate the two sets. We introduce VisDiff which first captions the images and prompts a language model to propose candidate descriptions then re-ranks these descriptions using CLIP. To evaluate VisDiff we collect VisDiffBench a dataset with 187 paired image sets with ground truth difference descriptions. We apply VisDiff to various domains such as comparing datasets (e.g. ImageNet vs. ImageNetV2) comparing classification models (e.g. zero-shot CLIP vs. supervised ResNet) characterizing differences between generative models (e.g. StableDiffusionV1 and V2) and discovering what makes images memorable. Using VisDiff we are able to find interesting and previously unknown differences in datasets and models demonstrating its utility in revealing nuanced insights.

Recovering a clear image from a single hazy image is an open inverse problem. Although significant research progress has been made most existing methods ignore the effect that downstream tasks play in promoting upstream dehazing. From the perspective of the haze generation mechanism there is a potential relationship between the depth information of the scene and the hazy image. Based on this we propose a dual-task collaborative mutual promotion framework to achieve the dehazing of a single image. This framework integrates depth estimation and dehazing by a dual-task interaction mechanism and achieves mutual enhancement of their performance. To realize the joint optimization of the two tasks an alternative implementation mechanism with the difference perception is developed. On the one hand the difference perception between the depth maps of the dehazing result and the ideal image is proposed to promote the dehazing network to pay attention to the non-ideal areas of the dehazing. On the other hand by improving the depth estimation performance in the difficult-to-recover areas of the hazy image the dehazing network can explicitly use the depth information of the hazy image to assist the clear image recovery. To promote the depth estimation we propose to use the difference between the dehazed image and the ground truth to guide the depth estimation network to focus on the dehazed unideal areas. It allows dehazing and depth estimation to leverage their strengths in a mutually reinforcing manner. Experimental results show that the proposed method can achieve better performance than that of the state-of-the-art approaches. The source code is released at https://github.com/zhoushen1/DCMPNet.

Vision Transformer (ViT) has emerged as a prominent architecture for various computer vision tasks. In ViT we divide the input image into patch tokens and process them through a stack of self-attention blocks. However unlike Convolutional Neural Network (CNN) ViT's simple architecture has no informative inductive bias (e.g. locality etc.). Due to this ViT requires a large amount of data for pre-training. Various data-efficient approaches (DeiT) have been proposed to train ViT on balanced datasets effectively. However limited literature discusses the use of ViT for datasets with long-tailed imbalances. In this work we introduce DeiT-LT to tackle the problem of training ViTs from scratch on long-tailed datasets. In DeiT-LT we introduce an efficient and effective way of distillation from CNN via distillation \texttt DIST token by using out-of-distribution images and re-weighting the distillation loss to enhance focus on tail classes. This leads to the learning of local CNN-like features in early ViT blocks improving generalization for tail classes. Further to mitigate overfitting we propose distilling from a flat CNN teacher which leads to learning low-rank generalizable features for DIST tokens across all ViT blocks. With the proposed DeiT-LT scheme the distillation DIST token becomes an expert on the tail classes and the classifier CLS token becomes an expert on the head classes. The experts help to effectively learn features corresponding to both the majority and minority classes using a distinct set of tokens within the same ViT architecture. We show the effectiveness of DeiT-LT for training ViT from scratch on datasets ranging from small-scale CIFAR-10 LT to large-scale iNaturalist-2018. Project Page: https://rangwani-harsh.github.io/DeiT-LT.

Inverse tone mapping (ITM) aims to reconstruct high dynamic range (HDR) radiance from low dynamic range (LDR) content. Although many deep image ITM methods can generate impressive results the field of video ITM is still to be explored. Processing video sequences by image ITM methods may cause temporal inconsistency. Besides they aren't able to exploit the potentially useful information in the temporal domain. In this paper we analyze the process of video filming and then propose a Global Sample and Local Propagate strategy to better find and utilize temporal clues. To better realize the proposed strategy we design a two-stage pipeline which includes modules named Incremental Clue Aggregation Module and Feature and Clue Propagation Module. They can align and fuse frames effectively under the condition of brightness changes and propagate features and temporal clues to all frames efficiently. Our temporal clues based video ITM method can recover realistic and temporal consistent results with high fidelity in over-exposed regions. Qualitative and quantitative experiments on public datasets show that the proposed method has significant advantages over existing methods.

In this paper we propose Image Downscaling Assessment by Rate-Distortion (IDA-RD) a novel measure to quantitatively evaluate image downscaling algorithms. In contrast to image-based methods that measure the quality of downscaled images ours is process-based that draws ideas from rate-distortion theory to measure the distortion incurred during downscaling. Our main idea is that downscaling and super-resolution (SR) can be viewed as the encoding and decoding processes in the rate-distortion model respectively and that a downscaling algorithm that preserves more details in the resulting low-resolution (LR) images should lead to less distorted high-resolution (HR) images in SR. In other words the distortion should increase as the downscaling algorithm deteriorates. However it is non-trivial to measure this distortion as it requires the SR algorithm to be blind and stochastic. Our key insight is that such requirements can be met by recent SR algorithms based on deep generative models that can find all matching HR images for a given LR image on their learned image manifolds. Extensive experimental results show the effectiveness of our IDA-RD measure.

Referring video segmentation relies on natural language expressions to identify and segment objects often emphasizing motion clues. Previous works treat a sentence as a whole and directly perform identification at the video-level mixing up static image-level cues with temporal motion cues. However image-level features cannot well comprehend motion cues in sentences and static cues are not crucial for temporal perception. In fact static cues can sometimes interfere with temporal perception by overshadowing motion cues. In this work we propose to decouple video-level referring expression understanding into static and motion perception with a specific emphasis on enhancing temporal comprehension. Firstly we introduce an expression-decoupling module to make static cues and motion cues perform their distinct role alleviating the issue of sentence embeddings overlooking motion cues. Secondly we propose a hierarchical motion perception module to capture temporal information effectively across varying timescales. Furthermore we employ contrastive learning to distinguish the motions of visually similar objects. These contributions yield state-of-the-art performance across five datasets including a remarkable 9.2% J&F improvement on the challenging MeViS dataset.

Medical vision language pre-training (VLP) has emerged as a frontier of research enabling zero-shot pathological recognition by comparing the query image with the textual descriptions for each disease. Due to the complex semantics of biomedical texts current methods struggle to align medical images with key pathological findings in unstructured reports. This leads to the misalignment with the target disease's textual representation. In this paper we introduce a novel VLP framework designed to dissect disease descriptions into their fundamental aspects leveraging prior knowledge about the visual manifestations of pathologies. This is achieved by consulting a large language model and medical experts. Integrating a Transformer module our approach aligns an input image with the diverse elements of a disease generating aspect-centric image representations. By consolidating the matches from each aspect we improve the compatibility between an image and its associated disease. Additionally capitalizing on the aspect-oriented representations we present a dual-head Transformer tailored to process known and unknown diseases optimizing the comprehensive detection efficacy. Conducting experiments on seven downstream datasets ours improves the accuracy of recent methods by up to 8.56% and 17.26% for seen and unseen categories respectively. Our code is released at https://github.com/HieuPhan33/MAVL.

Two-view correspondence learning has recently focused on considering the coherence and smoothness of the motion field between an image pair. Dominant schemes include controlling the complexity of the field function with regularization or smoothing the field with local filters but the former suffers from heavy computational burden and the latter fails to accommodate discontinuities in the case of large scene disparities. In this paper inspired by Fourier expansion we propose a novel network called DeMatch which decomposes the motion field to retain its main "low-frequency" and smooth part. This achieves implicit regularization with lower computational cost and generates piecewise smoothness naturally. Specifically we first decompose the rough motion field that is contaminated by false matches into several different sub-fields which are highly smooth and contain the main energy of the original field. Then with these smooth sub-fields we recover a cleaner motion field from which correct motion vectors are subsequently derived. We also design a special masked decomposition strategy to further mitigate the negative influence of false matches. All the mentioned processes are finally implemented in a discrete and learnable manner avoiding the difficulty of calculating real dense fields. Extensive experiments reveal that DeMatch outperforms state-of-the-art methods in multiple tasks and shows promising low computational usage and piecewise smoothness property. The code and trained models are publicly available at https://github.com/SuhZhang/DeMatch.

We demonstrate text as a strong cross-modal interface. Rather than relying on deep embeddings to connect image and language as the interface representation our approach represents an image as text from which we enjoy the interpretability and flexibility inherent to natural language. We employ an autoencoder that uses a pre-trained text-to-image diffusion model for decoding. The encoder is trained to transform an input image into text which is then fed into the fixed text-to-image diffusion decoder to reconstruct the original input a process we term De-Diffusion. Experiments validate both the precision and comprehensiveness of De-Diffusion text representing images such that it can be readily ingested by off-the-shelf text-to-image tools and LLMs for diverse multi-modal tasks. For example a single De-Diffusion model can generalize to provide transferable prompts for different text-to-image tools and also achieves a new state of the art on open-ended vision-language tasks by simply prompting large language models with few-shot examples. Project page: https://dediffusion.github.io/

The goal of multimodal alignment is to learn a single latent space that is shared between multimodal inputs. The most powerful models in this space have been trained using massive datasets of paired inputs and large-scale computational resources making them prohibitively expensive to train in many practical scenarios. We surmise that existing unimodal encoders pre-trained on large amounts of unimodal data should provide an effective bootstrap to create multimodal models from unimodal ones at much lower costs. We therefore propose FuseMix a multimodal augmentation scheme that operates on the latent spaces of arbitrary pre-trained unimodal encoders. Using FuseMix for multimodal alignment we achieve competitive performance - and in certain cases outperform state-of-the art methods - in both image-text and audio-text retrieval with orders of magnitude less compute and data: for example we outperform CLIP on the Flickr30K text-to-image retrieval task with ?600x fewer GPU days and ?80x fewer image-text pairs. Additionally we show how our method can be applied to convert pre-trained text-to-image generative models into audio-to-image ones. Code is available at: https://github.com/layer6ai-labs/fusemix.

Recently dataset distillation has paved the way towards efficient machine learning especially for image datasets. However the distillation for videos characterized by an exclusive temporal dimension remains an underexplored domain. In this work we provide the first systematic study of video distillation and introduce a taxonomy to categorize temporal compression. Our investigation reveals that the temporal information is usually not well learned during distillation and the temporal dimension of synthetic data contributes little. The observations motivate our unified framework of disentangling the dynamic and static information in the videos. It first distills the videos into still images as static memory and then compensates the dynamic and motion information with a learnable dynamic memory block. Our method achieves state-of-the-art on video datasets at different scales with notably smaller memory storage budget. Our code is available at https://github.com/yuz1wan/video_distillation.