CVPR 2024 Papers

Despite the remarkable success of Vision Transformers (ViT) across diverse fields in computer vision they have a clear drawback of expensive adaption cost for downstream tasks due to the increased scale. To address this Visual Prompt Tuning (VPT) incorporates learnable parameters in the input space of ViT. While freezing the ViT backbone and tuning only the prompts it exhibits superior performances to full fine-tuning. However despite the outstanding advantage we point out that VPT may lead to serious unfairness in downstream classification. Initially we investigate the causes of unfairness in VPT identifying the biasedly pre-trained ViT as a principal factor. Motivated by this observation we propose a Fair Visual Prompt Tuning (Fair-VPT) which removes biased information in the pre-trained ViT while adapting it to downstream classification tasks. To this end we categorize prompts into "cleaner prompts" and "target prompts". Based on this we encode the class token in two different ways by either masking or not masking the target prompts in the self-attention process. These encoded tokens are trained with distinct objective functions resulting in the inclusion of different information in the target and cleaner prompts. Moreover we introduce a disentanglement loss based on contrastive learning to further decorrelate them. In experiments across diverse benchmarks the proposed method demonstrates the most superior performance in terms of balanced classification accuracy and fairness.

Is vision good enough for language? Recent advancements in multimodal models primarily stem from the powerful reasoning abilities of large language models (LLMs). However the visual component typically depends only on the instance-level contrastive language-image pre-training (CLIP). Our research reveals that the visual capabilities in recent MultiModal LLMs (MLLMs) still exhibit systematic shortcomings. To understand the roots of these errors we explore the gap between the visual embedding space of CLIP and vision-only self-supervised learning. We identify "CLIP-blind pairs" - images that CLIP perceives as similar despite their clear visual differences. With these pairs we construct the Multimodal Visual Patterns (MMVP) benchmark. MMVP exposes areas where state-of-the-art systems including GPT-4V struggle with straightforward questions across nine basic visual patterns often providing incorrect answers and hallucinated explanations. We further evaluate various CLIP-based vision-and-language models and found a notable correlation between visual patterns that challenge CLIP models and those problematic for multimodal LLMs. As an initial effort to address these issues we propose a Mixture of Features (MoF) approach demonstrating that integrating vision self-supervised learning features with MLLMs can significantly enhance their visual grounding capabilities. Together our research suggests visual representation learning remains an open challenge and accurate visual grounding is crucial for future successful multimodal systems.

Although Multimodal Large Language Models (MLLMs) have demonstrated promising versatile capabilities their performance is still inferior to specialized models on downstream tasks which makes adaptation necessary to enhance their utility. However fine-tuning methods require independent training for every model leading to huge computation and memory overheads. In this paper we propose a novel setting where we aim to improve the performance of diverse MLLMs with a group of shared parameters optimized for a downstream task. To achieve this we propose Transferable Visual Prompting (TVP) a simple and effective approach to generate visual prompts that can transfer to different models and improve their performance on downstream tasks after trained on only one model. We introduce two strategies to address the issue of cross-model feature corruption of existing visual prompting methods and enhance the transferability of the learned prompts including 1) Feature Consistency Alignment: which imposes constraints to the prompted feature changes to maintain task-agnostic knowledge; 2) Task Semantics Enrichment: which encourages the prompted images to contain richer task-specific semantics with language guidance. We validate the effectiveness of TVP through extensive experiments with 6 modern MLLMs on a wide variety of tasks ranging from object recognition and counting to multimodal reasoning and hallucination correction.

Continual learning (CL) aims to empower models to learn new tasks without forgetting previously acquired knowledge. Most prior works concentrate on the techniques of architectures replay data regularization etc. However the category name of each class is largely neglected. Existing methods commonly utilize the one-hot labels and randomly initialize the classifier head. We argue that the scarce semantic information conveyed by the one-hot labels hampers the effective knowledge transfer across tasks. In this paper we revisit the role of the classifier head within the CL paradigm and replace the classifier with semantic knowledge from pretrained language models (PLMs). Specifically we use PLMs to generate semantic targets for each class which are frozen and serve as supervision signals during training. Such targets fully consider the semantic correlation between all classes across tasks. Empirical studies show that our approach mitigates forgetting by alleviating representation drifting and facilitating knowledge transfer across tasks. The proposed method is simple to implement and can seamlessly be plugged into existing methods with negligible adjustments. Extensive experiments based on eleven mainstream baselines demonstrate the effectiveness and generalizability of our approach to various protocols. For example under the class-incremental learning setting on ImageNet-100 our method significantly improves the Top-1 accuracy by 3.2% to 6.1% while reducing the forgetting rate by 2.6% to 13.1%.

Visual Instruction Tuning represents a novel learning paradigm involving the fine-tuning of pre-trained language models using task-specific instructions. This paradigm shows promising zero-shot results in various natural language processing tasks but is still unexplored in vision emotion understanding. In this work we focus on enhancing the model's proficiency in understanding and adhering to instructions related to emotional contexts. Initially we identify key visual clues critical to visual emotion recognition. Subsequently we introduce a novel GPT-assisted pipeline for generating emotion visual instruction data effectively addressing the scarcity of annotated instruction data in this domain. Expanding on the groundwork established by InstructBLIP our proposed EmoVIT architecture incorporates emotion-specific instruction data leveraging the powerful capabilities of Large Language Models to enhance performance. Through extensive experiments our model showcases its proficiency in emotion classification adeptness in affective reasoning and competence in comprehending humor. The comparative analysis provides a robust benchmark for Emotion Visual Instruction Tuning in the era of LLMs providing valuable insights and opening avenues for future exploration in this domain. Our code is available at https://github.com/aimmemotion/EmoVIT.

Event-based Vision Sensors (EVS) gain popularity in enhancing CMOS Image Sensor (CIS) video capture. Nonidealities of EVS such as pixel or readout latency can significantly influence the quality of the enhanced images and warrant dedicated consideration in the design of fusion algorithms. A novel approach for jointly computing deblurred rolling-shutter artifact corrected high-speed videos with frame rates up to 10000 FPS using inherently blurry rolling shutter CIS frames of 120 FPS to 150 FPS in conjunction with EVS data from a hybrid CIS-EVS sensor is presented. EVS pixel latency readout latency and the sensor's refractory period are explicitly incorporated into the measurement model. This inverse function problem is solved on a per-pixel manner using an optimization-based framework. The interpolated images are subsequently processed by a novel refinement network. The proposed method is evaluated using simulated and measured datasets under natural and controlled environments. Extensive experiments show reduced shadowing effect a 4 dB increment in PSNR and a 12% improvement in LPIPS score compared to state-of-the-art methods.

We propose a novel strategy ES3 for self-supervised learning of robust audio-visual speech representations from unlabeled talking face videos. While many recent approaches for this task primarily rely on guiding the learning process using the audio modality alone to capture information shared between audio and video we reframe the problem as the acquisition of shared unique (modality-specific) and synergistic speech information to address the inherent asymmetry between the modalities. Based on this formulation we propose a novel "evolving" strategy that progressively builds joint audio-visual speech representations that are strong for both uni-modal (audio & visual) and bi-modal (audio-visual) speech. First we leverage the more easily learnable audio modality to initialize audio and visual representations by capturing audio-unique and shared speech information. Next we incorporate video-unique speech information and bootstrap the audio-visual representations on top of the previously acquired shared knowledge. Finally we maximize the total audio-visual speech information including synergistic information to obtain robust and comprehensive representations. We implement ES3 as a simple Siamese framework and experiments on both English benchmarks and a newly contributed large-scale Mandarin dataset show its effectiveness. In particular on LRS2-BBC our smallest model is on par with SoTA models with only 1/2 parameters and 1/8 unlabeled data (223h).

Automatic web navigation aims to build a web agent that can follow language instructions to execute complex and diverse tasks on real-world websites. Existing work primarily takes HTML documents as input which define the contents and action spaces (i.e. actionable elements and operations) of webpages. Nevertheless HTML documents may not provide a clear task-related context for each element making it hard to select the right (sequence of) actions. In this paper we propose to contextualize HTML elements through their "dual views" in webpage screenshots: each HTML element has its corresponding bounding box and visual content in the screenshot. We build upon the insight---web developers tend to arrange task-related elements nearby on webpages to enhance user experiences---and propose to contextualize each element with its neighbor elements using both textual and visual features. The resulting representations of HTML elements are more informative for the agent to take action. We validate our method on the recently released Mind2Web dataset which features diverse navigation domains and tasks on real-world websites. Our method consistently outperforms the baseline in all the scenarios including cross-task cross-website and cross-domain ones.

We propose a controllable visual enhancer named DDBF which is based on cross-modal conditional adversarial learning and aims to dispel darkness and achieve better visible and infrared modalities fusion. Specifically a guided restoration module (GRM) is firstly designed to enhance weakened information in the low-light visible modality. The GRM utilizes the light-invariant high-contrast characteristics of the infrared modality as the central target distribution and constructs a multi-level conditional adversarial sample set to enable continuous controlled brightness enhancement of visible images. Then we develop an information fusion module (IFM) to integrate the advantageous features of the enhanced visible image and the infrared image. Thanks to customized explicit information preservation and hue fidelity constraints the IFM produces visually pleasing results with rich textures significant contrast and vivid colors. The brightened visible image and the final fused image compose the dual output of our DDBF to meet the diverse visual preferences of users. We evaluate DDBF on the public datasets achieving state-of-the-art performances of low-light enhancement and information integration that is available for both day and night scenarios. The experiments also demonstrate that our DDBF is effective in improving decision accuracy for object detection and semantic segmentation. Moreover we offer a user-friendly interface for the convenient application of our model. The code is publicly available at https://github.com/HaoZhang1018/DDBF.

Existing Siamese or transformer trackers commonly pose visual object tracking as a one-shot detection problem i.e. locating the target object in a single forward evaluation scheme. Despite the demonstrated success these trackers may easily drift towards distractors with similar appearance due to the single forward evaluation scheme lacking self-correction. To address this issue we cast visual tracking as a point set based denoising diffusion process and propose a novel generative learning based tracker dubbed DiffusionTrack. Our DiffusionTrack possesses two appealing properties: 1) It follows a novel noise-to-target tracking paradigm that leverages multiple denoising diffusion steps to localize the target in a dynamic searching manner per frame. 2) It models the diffusion process using a point set representation which can better handle appearance variations for more precise localization. One side benefit is that DiffusionTrack greatly simplifies the post-processing e.g. removing window penalty scheme. Without bells and whistles our DiffusionTrack achieves leading performance over the state-of-the-art trackers and runs in real-time. The code is in https://github.com/VISION-SJTU/DiffusionTrack.

Diffusion generative modeling has become a promising approach for learning robotic manipulation tasks from stochastic human demonstrations. In this paper we present Diffusion-EDFs a novel SE(3)-equivariant diffusion-based approach for visual robotic manipulation tasks. We show that our proposed method achieves remarkable data efficiency requiring only 5 to 10 human demonstrations for effective end-to-end training in less than an hour. Furthermore our benchmark experiments demonstrate that our approach has superior generalizability and robustness compared to state-of-the-art methods. Lastly we validate our methods with real hardware experiments.

Audio-visual saliency prediction can draw support from diverse modality complements but further performance enhancement is still challenged by customized architectures as well as task-specific loss functions. In recent studies denoising diffusion models have shown more promising in unifying task frameworks owing to their inherent ability of generalization. Following this motivation a novel Diffusion architecture for generalized audio-visual Saliency prediction (DiffSal) is proposed in this work which formulates the prediction problem as a conditional generative task of the saliency map by utilizing input audio and video as the conditions. Based on the spatio-temporal audio-visual features an extra network Saliency-UNet is designed to perform multi-modal attention modulation for progressive refinement of the ground-truth saliency map from the noisy map. Extensive experiments demonstrate that the proposed DiffSal can achieve excellent performance across six challenging audio-visual benchmarks with an average relative improvement of 6.3% over the previous state-of-the-art results by six metrics.

Advances in neural fields are enabling high-fidelity capture of the shape and appearance of dynamic 3D scenes. However their capabilities lag behind those offered by conventional representations such as 2D videos because of algorithmic challenges and the lack of large-scale multi-view real-world datasets. We address the dataset limitation with DiVa-360 a real-world 360? dynamic visual dataset that contains synchronized high-resolution and long-duration multi-view video sequences of table-scale scenes captured using a customized low-cost system with 53 cameras. It contains 21 object-centric sequences categorized by different motion types 25 intricate hand-object interaction sequences and 8 long-duration sequences for a total of 17.4 M image frames. In addition we provide foreground-background segmentation masks synchronized audio and text descriptions. We benchmark the state-of-the-art dynamic neural field methods on DiVa-360 and provide insights about existing methods and future challenges on long-duration neural field capture.

Learning joint and coordinated features across modalities is essential for many audio-visual tasks. Existing pre-training methods primarily focus on global information neglecting fine-grained features and positions leading to suboptimal performance in dense prediction tasks. To address this issue we take a further step towards region-aware audio-visual pre-training and propose CrossMAE which excels in Cross-modality interaction and region alignment. Specifically we devise two masked autoencoding (MAE) pretext tasks at both pixel and embedding levels namely Cross-Conditioned Reconstruction and Cross-Embedding Reconstruction. Taking the visual modality as an example (the same goes for audio) in Cross-Conditioned Reconstruction the visual modality reconstructs the input image pixels conditioned on audio Attentive Tokens. As for the more challenging Cross-Embedding Reconstruction unmasked visual tokens reconstruct complete audio features under the guidance of learnable queries implying positional information which effectively enhances the interaction between modalities and exploits fine-grained semantics. Experimental results demonstrate that CrossMAE achieves state-of-the-art performance not only in classification and retrieval but also in dense prediction tasks. Furthermore we dive into the mechanism of modal interaction and region alignment of CrossMAE highlighting the effectiveness of the proposed components.

Recently an audio-visual segmentation (AVS) task has been introduced aiming to group pixels with sounding objects within a given video. This task necessitates a first-ever audio-driven pixel-level understanding of the scene posing significant challenges. In this paper we propose an innovative audio-visual transformer framework termed COMBO an acronym for COoperation of Multi-order Bilateral relatiOns. For the first time our framework explores three types of bilateral entanglements within AVS: pixel entanglement modality entanglement and temporal entanglement. Regarding pixel entanglement we employ a Siam-Encoder Module (SEM) that leverages prior knowledge to generate more precise visual features from the foundational model. For modality entanglement we design a Bilateral-Fusion Module (BFM) enabling COMBO to align corresponding visual and auditory signals bi-directionally. As for temporal entanglement we introduce an innovative adaptive inter-frame consistency loss according to the inherent rules of temporal. Comprehensive experiments and ablation studies on AVSBench-object (84.7 mIoU on S4 59.2 mIou on MS3) and AVSBench-semantic (42.1 mIoU on AVSS) datasets demonstrate that COMBO surpasses previous state-of-the-art methods. Project page is available at https://yannqi.github.io/AVS-COMBO.

People are spending an enormous amount of time on digital devices through graphical user interfaces (GUIs) e.g. computer or smartphone screens. Large language models (LLMs) such as ChatGPT can assist people in tasks like writing emails but struggle to understand and interact with GUIs thus limiting their potential to increase automation levels. In this paper we introduce CogAgent an 18-billion-parameter visual language model (VLM) specializing in GUI understanding and navigation. By utilizing both low-resolution and high-resolution image encoders CogAgent supports input at a resolution of 1120*1120 enabling it to recognize tiny page elements and text. As a generalist visual language model CogAgent achieves the state of the art on five text-rich and four general VQA benchmarks including VQAv2 OK-VQA Text-VQA ST-VQA ChartQA infoVQA DocVQA MM-Vet and POPE. CogAgent using only screenshots as input outperforms LLM-based methods that consume extracted HTML text on both PC and Android GUI navigation tasks---Mind2Web and AITW advancing the state of the art. The model and codes are available at https://github.com/THUDM/CogVLM.

Existing open-vocabulary image segmentation methods require a fine-tuning step on mask labels and/or image-text datasets. Mask labels are labor-intensive which limits the number of categories in segmentation datasets. Consequently the vocabulary capacity of pre-trained VLMs is severely reduced after fine-tuning. However without fine-tuning VLMs trained under weak image-text supervision tend to make suboptimal mask predictions. To alleviate these issues we introduce a novel recurrent framework that progressively filters out irrelevant texts and enhances mask quality without training efforts. The recurrent unit is a two-stage segmenter built upon a frozen VLM. Thus our model retains the VLM's broad vocabulary space and equips it with segmentation ability. Experiments show that our method outperforms not only the training-free counterparts but also those fine-tuned with millions of data samples and sets the new state-of-the-art records for both zero-shot semantic and referring segmentation. Concretely we improve the current record by 28.8 16.0 and 6.9 mIoU on Pascal VOC COCO Object and Pascal Context.

In this work we investigate the potential of a large language model (LLM) to directly comprehend visual signals without the necessity of fine-tuning on multi-modal datasets. The foundational concept of our method views an image as a linguistic entity and translates it to a set of discrete words derived from the LLM's vocabulary. To achieve this we present the Vision-to-Language Tokenizer abbreviated as V2T Tokenizer which transforms an image into a "foreign language" with the combined aid of an encoder-decoder the LLM vocabulary and a CLIP model. With this innovative image encoding the LLM gains the ability not only for visual comprehension but also for image denoising and restoration in an auto-regressive fashion--crucially without any fine-tuning. We undertake rigorous experiments to validate our method encompassing understanding tasks like image recognition image captioning and visual question answering as well as image denoising tasks like inpainting outpainting deblurring and shift restoration. Code and models are available at https://github.com/zh460045050/V2L-Tokenizer.

Understanding how attention varies across individuals has significant scientific and societal impacts. However existing visual scanpath models treat attention uniformly neglecting individual differences. To bridge this gap this paper focuses on individualized scanpath prediction (ISP) a new attention modeling task that aims to accurately predict how different individuals shift their attention in diverse visual tasks. It proposes an ISP method featuring three novel technical components: (1) an observer encoder to characterize and integrate an observer's unique attention traits (2) an observer-centric feature integration approach that holistically combines visual features task guidance and observer-specific characteristics and (3) an adaptive fixation prioritization mechanism that refines scanpath predictions by dynamically prioritizing semantic feature maps based on individual observers' attention traits. These novel components allow scanpath models to effectively address the attention variations across different observers. Our method is generally applicable to different datasets model architectures and visual tasks offering a comprehensive tool for transforming general scanpath models into individualized ones. Comprehensive evaluations using value-based and ranking-based metrics verify the method's effectiveness and generalizability.