CVPR 2024 Papers

Consistency learning is a central strategy to tackle unlabeled data in semi-supervised medical image segmentation (SSMIS) which enforces the model to produce consistent predictions under the perturbation. However most current approaches solely focus on utilizing a specific single perturbation which can only cope with limited cases while employing multiple perturbations simultaneously is hard to guarantee the quality of consistency learning. In this paper we propose an Adaptive Bidirectional Displacement (ABD) approach to solve the above challenge. Specifically we first design a bidirectional patch displacement based on reliable prediction confidence for unlabeled data to generate new samples which can effectively suppress uncontrollable regions and still retain the influence of input perturbations. Meanwhile to enforce the model to learn the potentially uncontrollable content a bidirectional displacement operation with inverse confidence is proposed for the labeled images which generates samples with more unreliable information to facilitate model learning. Extensive experiments show that ABD achieves new state-of-the-art performances for SSMIS significantly improving different baselines. Source code is available at https://github.com/chy-upc/ABD.

Transformer-based approaches have achieved promising performance in image restoration tasks given their ability to model long-range dependencies which is crucial for recovering clear images. Though diverse efficient attention mechanism designs have addressed the intensive computations associated with using transformers they often involve redundant information and noisy interactions from irrelevant regions by considering all available tokens. In this work we propose an Adaptive Sparse Transformer (AST) to mitigate the noisy interactions of irrelevant areas and remove feature redundancy in both spatial and channel domains. AST comprises two core designs i.e. an Adaptive Sparse Self-Attention (ASSA) block and a Feature Refinement Feed-forward Network (FRFN). Specifically ASSA is adaptively computed using a two-branch paradigm where the sparse branch is introduced to filter out the negative impacts of low query-key matching scores for aggregating features while the dense one ensures sufficient information flow through the network for learning discriminative representations. Meanwhile FRFN employs an enhance-and-ease scheme to eliminate feature redundancy in channels enhancing the restoration of clear latent images. Experimental results on commonly used benchmarks have demonstrated the versatility and competitive performance of our method in several tasks including rain streak removal real haze removal and raindrop removal. The code and pre-trained models are available at https://github.com/joshyZhou/AST.

Although image super-resolution (SR) problem has experienced unprecedented restoration accuracy with deep neural networks it has yet limited versatile applications due to the substantial computational costs. Since different input images for SR face different restoration difficulties adapting computational costs based on the input image referred to as adaptive inference has emerged as a promising solution to compress SR networks. Specifically adapting the quantization bit-widths has successfully reduced the inference and memory cost without sacrificing the accuracy. However despite the benefits of the resultant adaptive network existing works rely on time-intensive quantization-aware training with full access to the original training pairs to learn the appropriate bit allocation policies which limits its ubiquitous usage. To this end we introduce the first on-the-fly adaptive quantization framework that accelerates the processing time from hours to seconds. We formulate the bit allocation problem with only two bit mapping modules: one to map the input image to the image-wise bit adaptation factor and one to obtain the layer-wise adaptation factors. These bit mappings are calibrated and fine-tuned using only a small number of calibration images. We achieve competitive performance with the previous adaptive quantization methods while the processing time is accelerated by x2000. Codes are available at https://github.com/Cheeun/AdaBM.

The pretraining-finetuning paradigm has gained popularity in various computer vision tasks. In this paradigm the emergence of active finetuning arises due to the abundance of large-scale data and costly annotation requirements. Active finetuning involves selecting a subset of data from an unlabeled pool for annotation facilitating subsequent finetuning. However the use of a limited number of training samples can lead to a biased distribution potentially resulting in model overfitting. In this paper we propose a new method called ActiveDC for the active finetuning tasks. Firstly we select samples for annotation by optimizing the distribution similarity between the subset to be selected and the entire unlabeled pool in continuous space. Secondly we calibrate the distribution of the selected samples by exploiting implicit category information in the unlabeled pool. The feature visualization provides an intuitive sense of the effectiveness of our approach to distribution calibration. We conducted extensive experiments on three image classification datasets with different sampling ratios. The results indicate that ActiveDC consistently outperforms the baseline performance in all image classification tasks. The improvement is particularly significant when the sampling ratio is low with performance gains of up to 10%. Our code will be released.

Recent advancements in machine learning have spotlighted the potential of hyperbolic spaces as they effectively learn hierarchical feature representations. While there has been progress in leveraging hyperbolic spaces in single-modality contexts its exploration in multimodal settings remains under explored. Some recent efforts have sought to transpose Euclidean multimodal learning techniques to hyperbolic spaces by adopting geodesic distance based contrastive losses. However we show both theoretically and empirically that such spatial proximity based contrastive loss significantly disrupts hierarchies in the latent space. To remedy this we advocate that the cross-modal representations should accept the inherent modality gap between text and images and introduce a novel approach to measure cross-modal similarity that does not enforce spatial proximity. Our approach show remarkable capabilities in preserving unimodal hierarchies while aligning the two modalities. Our experiments on a series of downstream tasks demonstrate that better latent structure emerges with our objective function while being superior in text-to-image and image-to-text retrieval tasks.

We present an approach to accelerate Neural Field training by efficiently selecting sampling locations. While Neural Fields have recently become popular it is often trained by uniformly sampling the training domain or through handcrafted heuristics. We show that improved convergence and final training quality can be achieved by a soft mining technique based on importance sampling: rather than either considering or ignoring a pixel completely we weigh the corresponding loss by a scalar. To implement our idea we use Langevin Monte-Carlo sampling. We show that by doing so regions with higher error are being selected more frequently leading to more than 2x improvement in convergence speed. The code and related resources for this study are publicly available at https://ubc-vision.github.io/nf-soft-mining/.

Recently pre-trained vision-language models (e.g. CLIP) have shown great potential in few-shot learning and attracted a lot of research interest. Although efforts have been made to improve few-shot ability of CLIP key factors on the effectiveness of existing methods have not been well studied limiting further exploration of CLIP's potential in few-shot learning. In this paper we first introduce a unified formulation to analyze CLIP-based few-shot learning methods from a perspective of logit bias which encourages us to learn an effective logit bias for further improving performance of CLIP-based few-shot learning methods. To this end we disassemble three key components involved in computation of logit bias (i.e. logit features logit predictor and logit fusion) and empirically analyze the effect on performance of few-shot classification. Based on analysis of key components this paper proposes a novel AMU-Tuning method to learn effective logit bias for CLIP-based few-shot classification. Specifically our AMU-Tuning predicts logit bias by exploiting the appropriate Auxiliary features which are fed into an efficient feature-initialized linear classifier with Multi-branch training. Finally an Uncertainty-based fusion is developed to incorporate logit bias into CLIP for few-shot classification. The experiments are conducted on several widely used benchmarks and the results show AMU-Tuning clearly outperforms its counterparts while achieving state-of-the-art performance of CLIP-based few-shot learning without bells and whistles.

A handful of visual foundation models (VFMs) have recently emerged as the backbones for numerous downstream tasks. VFMs like CLIP DINOv2 SAM are trained with distinct objectives exhibiting unique characteristics for various downstream tasks. We find that despite their conceptual differences these models can be effectively merged into a unified model through multi-teacher distillation. We name this approach AM-RADIO (Agglomerative Model -- Reduce All Domains Into One). This integrative approach not only surpasses the performance of individual teacher models but also amalgamates their distinctive features such as zero-shot vision-language comprehension detailed pixel-level understanding and open vocabulary segmentation capabilities. Additionally in pursuit of the most hardware-efficient backbone we evaluated numerous architectures in our multi-teacher distillation pipeline using the same training recipe. This led to the development of a novel architecture (E-RADIO) that exceeds the performance of its predecessors and is at least 6x faster than the teacher models at matched resolution. Our comprehensive benchmarking process covers downstream tasks including ImageNet classification semantic segmentation linear probing COCO object detection and integration into LLaVa-1.5.

This work presents Adaptive Local-then-Global Merging (ALGM) a token reduction method for semantic segmentation networks that use plain Vision Transformers. ALGM merges tokens in two stages: (1) In the first network layer it merges similar tokens within a small local window and (2) halfway through the network it merges similar tokens across the entire image. This is motivated by an analysis in which we found that in those situations tokens with a high cosine similarity can likely be merged without a drop in segmentation quality. With extensive experiments across multiple datasets and network configurations we show that ALGM not only significantly improves the throughput by up to 100% but can also enhance the mean IoU by up to +1.1 thereby achieving a better trade-off between segmentation quality and efficiency than existing methods. Moreover our approach is adaptive during inference meaning that the same model can be used for optimal efficiency or accuracy depending on the application. Code is available at https://tue-mps.github.io/ALGM.

Autonomous vehicle (AV) systems rely on robust perception models as a cornerstone of safety assurance. However objects encountered on the road exhibit a long-tailed distribution with rare or unseen categories posing challenges to a deployed perception model. This necessitates an expensive process of continuously curating and annotating data with significant human effort. We propose to leverage recent advances in vision-language and large language models to design an Automatic Data Engine (AIDE) that automatically identifies issues efficiently curates data improves the model through auto-labeling and verifies the model through generation of diverse scenarios. This process operates iteratively allowing for continuous self-improvement of the model. We further establish a benchmark for open-world detection on AV datasets to comprehensively evaluate various learning paradigms demonstrating our method's superior performance at a reduced cost.

Automatic radiology report generation using deep learning models has been recently explored and found promising. Neural decoders are commonly used for the report generation where irrelevant and unfaithful contents are unavoidable. The retrieval-based approach alleviates the limitation by identifying reports which are relevant to the input to assist the generation. To achieve clinically accurate report retrieval we make reference to clinicians' diagnostic steps of examining a radiology image where anatomical and diagnostic details are typically focused and propose a novel hierarchical visual concept representation called anatomy-aware hierarchical vision encoding (AHIVE). To learn AHIVE we first derive a methodology to extract hierarchical diagnostic descriptions from radiology reports and develop a CLIP-based framework for the model training. Also the hierarchical architecture of AHIVE is designed to support interactive report retrieval so that report revision made at one layer can be propagated to the subsequent ones to trigger other necessary revisions. We conduct extensive experiments and show that AHIVE can outperform the SOTA vision-language retrieval methods in terms of clinical accuracy by a large margin. We provide also a case study to illustrate how it enables interactive report retrieval.

What does learning to model relationships between strings teach Large Language Models (LLMs) about the visual world? We systematically evaluate LLMs' abilities to generate and recognize an assortment of visual concepts of increasing complexity and then demonstrate how a preliminary visual representation learning system can be trained using models of text. As language models lack the ability to consume or output visual information as pixels we use code to represent images in our study. Although LLM-generated images do not look like natural images results on image generation and the ability of models to correct these generated images indicate that precise modeling of strings can teach language models about numerous aspects of the visual world. Furthermore experiments on self-supervised visual representation learning utilizing images generated with text models highlight the potential to train vision models capable of making semantic assessments of natural images using just LLMs.

This paper presents a video inversion approach for zero-shot video editing which models the input video with low-rank representation during the inversion process. The existing video editing methods usually apply the typical 2D DDIM inversion or naive spatial-temporal DDIM inversion before editing which leverages time-varying representation for each frame to derive noisy latent. Unlike most existing approaches we propose a Spatial-Temporal Expectation-Maximization (STEM) inversion which formulates the dense video feature under an expectation-maximization manner and iteratively estimates a more compact basis set to represent the whole video. Each frame applies the fixed and global representation for inversion which is more friendly for temporal consistency during reconstruction and editing. Extensive qualitative and quantitative experiments demonstrate that our STEM inversion can achieve consistent improvement on two state-of-the-art video editing methods. Project page: https://stem-inv.github.io/page/.

Large-scale diffusion generative models are greatly simplifying image video and 3D asset creation from user provided text prompts and images. However the challenging problem of text-to-4D dynamic 3D scene generation with diffusion guidance remains largely unexplored. We propose Dream-in-4D which features a novel two-stage approach for text-to-4D synthesis leveraging (1) 3D and 2D diffusion guidance to effectively learn a high-quality static 3D asset in the first stage; (2) a deformable neural radiance field that explicitly disentangles the learned static asset from its deformation preserving quality during motion learning; and (3) a multi-resolution feature grid for the deformation field with a displacement total variation loss to effectively learn motion with video diffusion guidance in the second stage. Through a user preference study we demonstrate that our approach significantly advances image and motion quality 3D consistency and text fidelity for text-to-4D generation compared to baseline approaches. Thanks to its motion-disentangled representation Dream-in-4D can also be easily adapted for controllable generation where appearance is defined by one or multiple images without the need to modify the motion learning stage. Thus our method offers for the first time a unified approach for text-to-4D image-to-4D and personalized 4D generation tasks.

We present a novel theory that establishes the relationship between light transport in visible and thermal infrared and heat transport in solids. We show that heat generated due to light absorption can be estimated by modeling heat transport using a thermal camera. For situations where heat conduction is negligible we analytically solve the heat transport equation to derive a simple expression relating the change in thermal image intensity to the absorbed light intensity and heat capacity of the material. Next we prove that intrinsic image decomposition for Lambertian scenes becomes a well-posed problem if one has access to the absorbed light. Our theory generalizes to arbitrary shapes and unstructured illumination. Our theory is based on applying energy conservation principle at each pixel independently. We validate our theory using real-world experiments on diffuse objects made of different materials that exhibit both direct and global components (inter-reflections) of light transport under unknown complex lighting.

Generalization to new domains not seen during training is one of the long-standing challenges in deploying neural networks in real-world applications. Existing generalization techniques either necessitate external images for augmentation and/or aim at learning invariant representations by imposing various alignment constraints. Large-scale pretraining has recently shown promising generalization capabilities along with the potential of binding different modalities. For instance the advent of vision-language models like CLIP has opened the doorway for vision models to exploit the textual modality. In this paper we introduce a simple framework for generalizing semantic segmentation networks by employing language as the source of randomization. Our recipe comprises three key ingredients: (i) the preservation of the intrinsic CLIP robustness through minimal fine-tuning (ii) language-driven local style augmentation and (iii) randomization by locally mixing the source and augmented styles during training. Extensive experiments report state-of-the-art results on various generalization benchmarks.

Understanding long real-world videos requires modeling of long-range visual dependencies. To this end we explore video-first architectures building on the common paradigm of transferring large-scale image--text models to video via shallow temporal fusion. However we expose two limitations to the approach: (1) decreased spatial capabilities likely due to poor video--language alignment in standard video datasets and (2) higher memory consumption bottlenecking the number of frames that can be processed. To mitigate the memory bottleneck we systematically analyze the memory/accuracy trade-off of various efficient methods: factorized attention parameter-efficient image-to-video adaptation input masking and multi-resolution patchification. Surprisingly simply masking large portions of the video (up to 75%) during contrastive pre-training proves to be one of the most robust ways to scale encoders to videos up to 4.3 minutes at 1 FPS. Our simple approach for training long video-to-text models which scales to 1B parameters does not add new architectural complexity and is able to outperform the popular paradigm of using much larger LLMs as an information aggregator over segment-based information on benchmarks with long-range temporal dependencies (YouCook2 EgoSchema).

Curation methods for massive vision-language datasets trade off between dataset size and quality. However even the highest quality of available curated captions are far too short to capture the rich visual detail in an image. To show the value of dense and highly-aligned image-text pairs we collect the Densely Captioned Images (DCI) dataset containing 8012 natural images human-annotated with mask-aligned descriptions averaging above 1000 words each. With precise and reliable captions associated with specific parts of an image we can evaluate vision-language models' (VLMs) understanding of image content with a novel task that matches each caption with its corresponding subcrop. As current models are often limited to 77 text tokens we also introduce a summarized version (sDCI) in which each caption length is limited. We show that modern techniques that make progress on standard benchmarks do not correspond with significant improvement on our sDCI based benchmark. Lastly we finetune CLIP using sDCI and show significant improvements over the baseline despite a small training set. By releasing the first human annotated dense image captioning dataset we hope to enable the development of new benchmarks or fine-tuning recipes for the next generation of VLMs to come.

Extensive advancements have been made in person ReID through the mining of semantic information. Nevertheless existing methods that utilize semantic-parts from a single image modality do not explicitly achieve this goal. Whiteness the impressive capabilities in multimodal understanding of Vision Language Foundation Model CLIP a recent two-stage CLIP-based method employs automated prompt engineering to obtain specific textual labels for classifying pedestrians. However we note that the predefined soft prompts may be inadequate in expressing the entire visual context and struggle to generalize to unseen classes. This paper presents an end-to-end Prompt-driven Semantic Guidance (PromptSG) framework that harnesses the rich semantics inherent in CLIP. Specifically we guide the model to attend to regions that are semantically faithful to the prompt. To provide personalized language descriptions for specific individuals we propose learning pseudo tokens that represent specific visual contexts. This design not only facilitates learning fine-grained attribute information but also can inherently leverage language prompts during inference. Without requiring additional labeling efforts our PromptSG achieves state-of-the-art by over 10% on MSMT17 and nearly 5% on the Market-1501 benchmark.