CVPR 2024 Papers

Large language models (LLMs) are fine-tuned using human comparison data with Reinforcement Learning from Human Feedback (RLHF) methods to make them better aligned with users' preferences. In contrast to LLMs human preference learning has not been widely explored in text-to-image diffusion models; the best existing approach is to fine-tune a pretrained model using carefully curated high quality images and captions to improve visual appeal and text alignment. We propose Diffusion-DPO a method to align diffusion models to human preferences by directly optimizing on human comparison data. Diffusion-DPO is adapted from the recently developed Direct Preference Optimization (DPO) a simpler alternative to RLHF which directly optimizes a policy that best satisfies human preferences under a classification objective. We re-formulate DPO to account for a diffusion model notion of likelihood utilizing the evidence lower bound to derive a differentiable objective. Using the Pick-a-pic dataset of 851K crowdsourced pairwise preferences we fine-tune the base model of the state-of-the-art Stable Diffusion XL (SDXL)-1.0 model with Diffusion-DPO. Our fine-tuned base model significantly outperforms both base SDXL-1.0 and the larger SDXL-1.0 model consisting of an additional refinement model in human evaluation improving visual appeal and prompt alignment. We also develop a variant that uses AI feedback and has comparable performance to training on human preferences opening the door for scaling of diffusion model alignment methods.

Photometric stereo faces challenges from non-Lambertian reflectance in real-world scenarios. Systematically measuring the reliability of photometric stereo methods in handling such complex reflectance necessitates a real-world dataset with quantitatively controlled reflectances. This paper introduces DiLiGenRT the first real-world dataset for evaluating photometric stereo methods under quantified reflectances by manufacturing 54 hemispheres with varying degrees of two reflectance properties: Roughness and Translucency. Unlike qualitative and semantic labels such as diffuse and specular that have been used in previous datasets our quantified dataset allows comprehensive and systematic benchmark evaluations. In addition it facilitates selecting best-fit photometric stereo methods based on the quantitative reflectance properties. Our dataset and benchmark results are available at https://photometricstereo.github.io/diligentrt.html.

While deep learning has led to huge progress in complex image classification tasks like ImageNet unexpected failure modes e.g. via spurious features call into question how reliably these classifiers work in the wild. Furthermore for safety-critical tasks the black-box nature of their decisions is problematic and explanations or at least methods which make decisions plausible are needed urgently. In this paper we address these problems by generating images that optimize a classifier-derived objective using a framework for guided image generation. We analyze the decisions of image classifiers by visual counterfactual explanations (VCEs) detection of systematic mistakes by analyzing images where classifiers maximally disagree and visualization of neurons and spurious features. In this way we validate existing observations e.g. the shape bias of adversarially robust models as well as novel failure modes e.g. systematic errors of zero-shot CLIP classifiers. Moreover our VCEs outperform previous work while being more versatile.

While standard Empirical Risk Minimization (ERM) training is proven effective for image classification on in-distribution data it fails to perform well on out-of-distribution samples. One of the main sources of distribution shift for image classification is the compositional nature of images. Specifically in addition to the main object or component(s) determining the label some other image components usually exist which may lead to the shift of input distribution between train and test environments. More importantly these components may have spurious correlations with the label. To address this issue we propose Decompose-and-Compose (DaC) which improves robustness to correlation shift by a compositional approach based on combining elements of images. Based on our observations models trained with ERM usually highly attend to either the causal components or the components having a high spurious correlation with the label (especially in datapoints on which models have a high confidence). In fact according to the amount of spurious correlation and the easiness of classification based on the causal or non-causal components the model usually attends to one of these more (on samples with high confidence). Following this we first try to identify the causal components of images using class activation maps of models trained with ERM. Afterward we intervene on images by combining them and retraining the model on the augmented data including the counterfactual ones. This work proposes a group-balancing method by intervening on images without requiring group labels or information regarding the spurious features during training. The method has an overall better worst group accuracy compared to previous methods with the same amount of supervision on the group labels in correlation shift. Our code is available at https://github.com/fhn98/DaC.

Federated Learning (FL) enables collaborative model training while preserving the privacy of raw data. A challenge in this framework is the fair and efficient valuation of data which is crucial for incentivizing clients to contribute high-quality data in the FL task. In scenarios involving numerous data clients within FL it is often the case that only a subset of clients and datasets are pertinent to a specific learning task while others might have either a negative or negligible impact on the model training process. This paper introduces a novel privacy-preserving method for evaluating client contributions and selecting relevant datasets without a pre-specified training algorithm in an FL task. Our proposed approach \texttt FedBary utilizes Wasserstein distance within the federated context offering a new solution for data valuation in the FL framework. This method ensures transparent data valuation and efficient computation of the Wasserstein barycenter and reduces the dependence on validation datasets. Through extensive empirical experiments and theoretical analyses we demonstrate the advantages of this data valuation method as a promising avenue for FL research.

Implicit neural representations for video (NeRV) have recently become a novel way for high-quality video representation. However existing works employ a single network to represent the entire video which implicitly confuse static and dynamic information. This leads to an inability to effectively compress the redundant static information and lack the explicitly modeling of global temporal-coherent dynamic details. To solve above problems we propose DS-NeRV which decomposes videos into sparse learnable static codes and dynamic codes without the need for explicit optical flow or residual supervision. By setting different sampling rates for two codes and applying weighted sum and interpolation sampling methods DS-NeRV efficiently utilizes redundant static information while maintaining high-frequency details. Additionally we design a cross-channel attention-based (CCA) fusion module to efficiently fuse these two codes for frame decoding. Our approach achieves a high quality reconstruction of 31.2 PSNR with only 0.35M parameters thanks to separate static and dynamic codes representation and outperforms existing NeRV methods in many downstream tasks. Our project website is at https://haoyan14.github.io/DS-NeRV.

Since the widespread availability of cameras black-and-white (BW)photography has been a popular choice for artistic and aesthetic expression. It highlights the main subject in varying tones of gray creating various effects such as drama and contrast. However producing BW photography often demands high-end cameras or photographic editing from experts. Even the experts prefer different styles depending on the subject or even the same subject when taking grayscale photos or converting color images to BW. It is thus questionable which approach is better. To imitate the artistic values of decolorized images this paper introduces a deep metric learning framework with a novel subject-style specified proxy and a large-scale BW dataset. Our proxy-based decolorization utilizes a hierarchical proxy-based loss and a hierarchical bilateral grid network to mimic the experts' retouching scheme. The proxy-based loss captures both expert-discriminative and classsharing characteristics while the hierarchical bilateral grid network enables imitating spatially-variant retouching by considering both global and local scene contexts. Our dataset including color and BW images edited by three experts demonstrates the scalability of our method which can be further enhanced by constructing additional proxies from any set of BW photos like Internet downloaded figures. Our Experiments show that our framework successfully produce visually-pleasing BW images from color ones as evaluated by user preference with respect to artistry and aesthetics.

Decentralized federated learning (DFL) facilitates collaborative model training across multiple connected clients without a central coordination server thereby avoiding the single point of failure in traditional centralized federated learning (CFL). However DFL exhibits heightened susceptibility to Byzantine attacks owing to the lack of a responsible central server. Furthermore a benign client in DFL may be dominated by Byzantine clients (more than half of its neighbors are malicious) posing significant challenges for robust model training. In this work we propose DFL-Dual a novel Byzantine-robust DFL method through dual-domain client clustering and trust bootstrapping. Specifically we first propose to leverage both data-domain and model-domain distance metrics to identify client discrepancies. Then we design a trust evaluation mechanism centered on benign clients which enables them to evaluate their neighbors. Building upon the dual-domain distance metric and trust evaluation mechanism we further develop a two-stage clustering and trust bootstrapping technique to exclude Byzantine clients from local model aggregation. We extensively evaluate the proposed DFL-Dual method through rigorous experimentation demonstrating its remarkable performance superiority over existing robust CFL and DFL schemes.

Modularity plays a crucial role in the development and maintenance of complex systems. While end-to-end text spotting efficiently mitigates the issues of error accumulation and sub-optimal performance seen in traditional two-step methodologies the two-step methods continue to be favored in many competitions and practical settings due to their superior modularity. In this paper we introduce Bridging Text Spotting a novel approach that resolves the error accumulation and suboptimal performance issues in two-step methods while retaining modularity. To achieve this we adopt a well-trained detector and recognizer that are developed and trained independently and then lock their parameters to preserve their already acquired capabilities. Subsequently we introduce a Bridge that connects the locked detector and recognizer through a zero-initialized neural network. This zero-initialized neural network initialized with weights set to zeros ensures seamless integration of the large receptive field features in detection into the locked recognizer. Furthermore since the fixed detector and recognizer cannot naturally acquire end-to-end optimization features we adopt the Adapter to facilitate their efficient learning of these features. We demonstrate the effectiveness of the proposed method through extensive experiments: Connecting the latest detector and recognizer through Bridging Text Spotting we achieved an accuracy of 83.3% on Total-Text 69.8% on CTW1500 and 89.5% on ICDAR 2015. The code is available at https://github.com/mxin262/Bridging-Text-Spotting.

Most existing dynamic or runtime channel pruning methods have to store all weights to achieve efficient inference which brings extra storage costs. Static pruning methods can reduce storage costs directly but their performance is limited by using a fixed sub-network to approximate the original model. Most existing pruning works suffer from these drawbacks because they were designed to only conduct either static or dynamic pruning. In this paper we propose a novel method to solve both efficiency and storage challenges via simultaneously conducting dynamic and static channel pruning for convolutional neural networks. We propose a new bi-level optimization based model to naturally integrate the static and dynamic channel pruning. By doing so our method enjoys benefits from both sides and the disadvantages of dynamic and static pruning are reduced. After pruning we permanently remove redundant parameters and then finetune the model with dynamic flexibility. Experimental results on CIFAR-10 and ImageNet datasets suggest that our method can achieve state-of-the-art performance compared to existing dynamic and static channel pruning methods.

Learning from seen attribute-object pairs to generalize to unseen compositions has been studied extensively in Compositional Zero-Shot Learning (CZSL). However CZSL setup is still limited to seen attributes and objects and cannot generalize to unseen concepts and their compositions. To overcome this limitation we propose a new task Open Vocabulary-Compositional Zero-shot Learning (OV-CZSL) where unseen attributes objects and unseen compositions are evaluated. To show that OV-CZSL is a challenging yet solvable problem we propose three new benchmarks based on existing datasets MIT-States C-GQA and VAW-CZSL along with new baselines and evaluation setup. We use language embeddings and external vocabulary with our novel neighborhood expansion loss to allow any method to learn semantic correlations between seen and unseen primitives.

This paper studies amodal image segmentation: predicting entire object segmentation masks including both visible and invisible (occluded) parts. In previous work the amodal segmentation ground truth on real images is usually predicted by manual annotaton and thus is subjective. In contrast we use 3D data to establish an automatic pipeline to determine authentic ground truth amodal masks for partially occluded objects in real images. This pipeline is used to construct an amodal completion evaluation benchmark MP3D-Amodal consisting of a variety of object categories and labels. To better handle the amodal completion task in the wild we explore two architecture variants: a two-stage model that first infers the occluder followed by amodal mask completion; and a one-stage model that exploits the representation power of Stable Diffusion for amodal segmentation across many categories. Without bells and whistles our method achieves a new state-of-the-art performance on Amodal segmentation datasets that cover a large variety of objects including COCOA and our new MP3D-Amodal dataset. The dataset model and code are available at https://www. robots.ox.ac.uk/ vgg/research/amodal/.

Scale-ambiguity in 3D scene dimensions leads to magnitude-ambiguity of volumetric densities in neural radiance fields i.e. the densities double when scene size is halved and vice versa. We call this property alpha invariance. For NeRFs to better maintain alpha invariance we recommend 1) parameterizing both distance and volume densities in log space and 2) a discretization-agnostic initialization strategy to guarantee high ray transmittance. We revisit a few popular radiance field models and find that these systems use various heuristics to deal with issues arising from scene scaling. We test their behaviors and show our recipe to be more robust.

Vision foundation models have been explored recently to build general-purpose vision systems. However predominant paradigms driven by casting instance-level tasks as an object-word alignment bring heavy cross-modality interaction which is not effective in prompting object detection and visual grounding. Another line of work that focuses on pixel-level tasks often encounters a large annotation gap of things and stuff and suffers from mutual interference between foreground-object and background-class segmentation. In stark contrast to the prevailing methods we present APE a universal visual perception model for aligning and prompting everything all at once in an image to perform diverse tasks i.e. detection segmentation and grounding as an instance-level sentence-object matching paradigm. Specifically APE advances the convergence of detection and grounding by reformulating language-guided grounding as open-vocabulary detection which efficiently scales up model prompting to thousands of category vocabularies and region descriptions while maintaining the effectiveness of cross-modality fusion. To bridge the granularity gap of different pixel-level tasks APE equalizes semantic and panoptic segmentation to proxy instance learning by considering any isolated regions as individual instances. APE aligns vision and language representation on broad data with natural and challenging characteristics all at once without task-specific fine-tuning. The extensive experiments on over 160 datasets demonstrate that with only one-suit of weights APE outperforms (or is on par with) the state-of-the-art models proving that an effective yet universal perception for anything aligning and prompting is indeed feasible. Codes and trained models are released at https://github.com/shenyunhang/APE.

Despite the recent progress made in Video Question-Answering (VideoQA) these methods typically function as black-boxes making it difficult to understand their reasoning processes and perform consistent compositional reasoning. To address these challenges we propose a model-agnostic Video Alignment and Answer Aggregation (VA3) framework which is capable of enhancing both compositional consistency and accuracy of existing VidQA methods by integrating video aligner and answer aggregator modules. The video aligner hierarchically selects the relevant video clips based on the question while the answer aggregator deduces the answer to the question based on its sub-questions with compositional consistency ensured by the information flow along the question decompose graph and the contrastive learning strategy. We evaluate our framework on three settings of the AGQA-Decomp dataset with three baseline methods and propose new metrics to measure the compositional consistency of VidQA methods more comprehensively. Moreover we propose a large language model (LLM) based automatic question decompose pipeline to apply our framework on any VidQA data. We extend MSVD and NExT-QA datasets with it to evaluate such scheme and our VA3 framework on broader scenarios. Extensive experiments show that our framework improves both compositional consistency and accuracy of existing methods leading to more interpretable models in real-world applications.

Expressive human pose and shape estimation (a.k.a. 3D whole-body mesh recovery) involves the human body hand and expression estimation. Most existing methods have tackled this task in a two-stage manner first detecting the human body part with an off-the-shelf detection model and then inferring the different human body parts individually. Despite the impressive results achieved these methods suffer from 1) loss of valuable contextual information via cropping 2) introducing distractions and 3) lacking inter-association among different persons and body parts inevitably causing performance degradation especially for crowded scenes. To address these issues we introduce a novel all-in-one-stage framework AiOS for multiple expressive human pose and shape recovery without an additional human detection step. Specifically our method is built upon DETR which treats multi-person whole-body mesh recovery task as a progressive set prediction problem with various sequential detection. We devise the decoder tokens and extend them to our task. Specifically we first employ a human token to probe a human location in the image and encode global features for each instance which provides a coarse location for the later transformer block. Then we introduce a joint-related token to probe the human joint in the image and encoder a fine-grained local feature which collaborates with the global feature to regress the whole-body mesh. This straightforward but effective model outperforms previous state-of-the-art methods by a 9 reduction in NMVE on AGORA a 30 reduction in PVE on EHF a 10 reduction in PVE on ARCTIC and a 3 reduction in PVE on EgoBody.

Adversarial distillation (AD) is a highly effective method for enhancing the robustness of small models. Contrary to expectations a high-performing teacher model does not always result in a more robust student model. This is due to two main reasons. First when there are significant differences in predictions between the teacher model and the student model exact matching of predicted values using KL divergence interferes with training leading to poor performance of existing methods. Second matching solely based on the output prevents the student model from fully understanding the behavior of the teacher model. To address these challenges this paper proposes a novel AD method named SmaraAD. During the training process we facilitate the student model in better understanding the teacher model's behavior by aligning the attribution region that the student model focuses on with that of the teacher model. Concurrently we relax the condition of exact matching in KL divergence and replace it with a more flexible matching criterion thereby enhancing the model's robustness. Extensive experiments substantiate the effectiveness of our method in improving the robustness of small models outperforming previous SOTA methods.

Keypoints used for image matching often include an estimate of the feature scale and orientation. While recent work has demonstrated the advantages of using feature scales and orientations for relative pose estimation relatively little work has considered their use for absolute pose estimation. We introduce minimal solutions for absolute pose from two oriented feature correspondences in the general case or one scaled and oriented correspondence given a known vertical direction. Nowadays assuming a known direction is not particularly restrictive as modern consumer devices such as smartphones or drones are equipped with Inertial Measurement Units (IMU) that provide the gravity direction by default. Compared to traditional absolute pose methods requiring three point correspondences our solvers need a smaller minimal sample reducing the cost and complexity of robust estimation. Evaluations on large-scale and public real datasets demonstrate the advantage of our methods for fast and accurate localization in challenging conditions. Code is available at https://github.com/danini/absolute-pose-from-oriented-and-scaled-features .