CVPR 2023 Papers

Deep hashing is an appealing approach for large-scale image retrieval. Most existing supervised deep hashing methods learn hash functions using pairwise or triple image similarities in randomly sampled mini-batches. They suffer from low training efficiency, insufficient coverage of data distribution, and pair imbalance problems. Recently, central similarity quantization (CSQ) attacks the above problems by using "hash centers" as a global similarity metric, which encourages the hash codes of similar images to approach their common hash center and distance themselves from other hash centers. Although achieving SOTA retrieval performance, CSQ falls short of a worst-case guarantee on the minimal distance between its constructed hash centers, i.e. the hash centers can be arbitrarily close. This paper presents an optimization method that finds hash centers with a constraint on the minimal distance between any pair of hash centers, which is non-trivial due to the non-convex nature of the problem. More importantly, we adopt the Gilbert-Varshamov bound from coding theory, which helps us to obtain a large minimal distance while ensuring the empirical feasibility of our optimization approach. With these clearly-separated hash centers, each is assigned to one image class, we propose several effective loss functions to train deep hashing networks. Extensive experiments on three datasets for image retrieval demonstrate that the proposed method achieves superior retrieval performance over the state-of-the-art deep hashing methods.

Learning a generalizable and comprehensive similarity metric to depict the semantic discrepancies between images is the foundation of many computer vision tasks. While existing methods approach this goal by learning an ensemble of embeddings with diverse objectives, the backbone network still receives a mix of all the training signals. Differently, we propose a deep factorized metric learning method (DFML) to factorize the training signal and employ different samples to train various components of the backbone network. We factorize the network to different sub-blocks and devise a learnable router to adaptively allocate the training samples to each sub-block with the objective to capture the most information. The metric model trained by DFML captures different characteristics with different sub-blocks and constitutes a generalizable metric when using all the sub-blocks. The proposed DFML achieves state-of-the-art performance on all three benchmarks for deep metric learning including CUB-200-2011, Cars196, and Stanford Online Products. We also generalize DFML to the image classification task on ImageNet-1K and observe consistent improvement in accuracy/computation trade-off. Specifically, we improve the performance of ViT-B on ImageNet (+0.2% accuracy) with less computation load (-24% FLOPs).

Robust and accurate geometric understanding against adverse weather conditions is one top prioritized conditions to achieve a high-level autonomy of self-driving cars. However, autonomous driving algorithms relying on the visible spectrum band are easily impacted by weather and lighting conditions. A long-wave infrared camera, also known as a thermal imaging camera, is a potential rescue to achieve high-level robustness. However, the missing necessities are the well-established large-scale dataset and public benchmark results. To this end, in this paper, we first built a large-scale Multi-Spectral Stereo (MS^2) dataset, including stereo RGB, stereo NIR, stereo thermal, and stereo LiDAR data along with GNSS/IMU information. The collected dataset provides about 195K synchronized data pairs taken from city, residential, road, campus, and suburban areas in the morning, daytime, and nighttime under clear-sky, cloudy, and rainy conditions. Secondly, we conduct an exhaustive validation process of monocular and stereo depth estimation algorithms designed on visible spectrum bands to benchmark their performance in the thermal image domain. Lastly, we propose a unified depth network that effectively bridges monocular depth and stereo depth tasks from a conditional random field approach perspective. Our dataset and source code are available at https://github.com/UkcheolShin/MS2-MultiSpectralStereoDataset.

Semantic editing of images is the fundamental goal of computer vision. Although deep learning methods, such as generative adversarial networks (GANs), are capable of producing high-quality images, they often do not have an inherent way of editing generated images semantically. Recent studies have investigated a way of manipulating the latent variable to determine the images to be generated. However, methods that assume linear semantic arithmetic have certain limitations in terms of the quality of image editing, whereas methods that discover nonlinear semantic pathways provide non-commutative editing, which is inconsistent when applied in different orders. This study proposes a novel method called deep curvilinear editing (DeCurvEd) to determine semantic commuting vector fields on the latent space. We theoretically demonstrate that owing to commutativity, the editing of multiple attributes depends only on the quantities and not on the order. Furthermore, we experimentally demonstrate that compared to previous methods, the nonlinear and commutative nature of DeCurvEd provides higher-quality editing.

The ability of scale-equivariance processing blocks plays a central role in arbitrary-scale image super-resolution tasks. Inspired by this crucial observation, this work proposes two novel scale-equivariant modules within a transformer-style framework to enhance arbitrary-scale image super-resolution (ASISR) performance, especially in high upsampling rate image extrapolation. In the feature extraction phase, we design a plug-in module called Adaptive Feature Extractor, which injects explicit scale information in frequency-expanded encoding, thus achieving scale-adaption in representation learning. In the upsampling phase, a learnable Neural Kriging upsampling operator is introduced, which simultaneously encodes both relative distance (i.e., scale-aware) information as well as feature similarity (i.e., with priori learned from training data) in a bilateral manner, providing scale-encoded spatial feature fusion. The above operators are easily plugged into multiple stages of a SR network, and a recent emerging pre-training strategy is also adopted to impulse the model's performance further. Extensive experimental results have demonstrated the outstanding scale-equivariance capability offered by the proposed operators and our learning framework, with much better results than previous SOTAs at arbitrary scales for SR. Our code is available at https://github.com/neuralchen/EQSR.

The images captured under improper exposure conditions often suffer from contrast degradation and detail distortion. Contrast degradation will destroy the statistical properties of low-frequency components, while detail distortion will disturb the structural properties of high-frequency components, leading to the low-frequency and high-frequency components being mixed and inseparable. This will limit the statistical and structural modeling capacity for exposure correction. To address this issue, this paper proposes to decouple the contrast enhancement and detail restoration within each convolution process. It is based on the observation that, in the local regions covered by convolution kernels, the feature response of low-/high-frequency can be decoupled by addition/difference operation. To this end, we inject the addition/difference operation into the convolution process and devise a Contrast Aware (CA) unit and a Detail Aware (DA) unit to facilitate the statistical and structural regularities modeling. The proposed CA and DA can be plugged into existing CNN-based exposure correction networks to substitute the Traditional Convolution (TConv) to improve the performance. Furthermore, to maintain the computational costs of the network without changing, we aggregate two units into a single TConv kernel using structural re-parameterization. Evaluations of nine methods and five benchmark datasets demonstrate that our proposed method can comprehensively improve the performance of existing methods without introducing extra computational costs compared with the original networks. The codes will be publicly available.

Human multimodal emotion recognition (MER) aims to perceive human emotions via language, visual and acoustic modalities. Despite the impressive performance of previous MER approaches, the inherent multimodal heterogeneities still haunt and the contribution of different modalities varies significantly. In this work, we mitigate this issue by proposing a decoupled multimodal distillation (DMD) approach that facilitates flexible and adaptive crossmodal knowledge distillation, aiming to enhance the discriminative features of each modality. Specially, the representation of each modality is decoupled into two parts, i.e., modality-irrelevant/-exclusive spaces, in a self-regression manner. DMD utilizes a graph distillation unit (GD-Unit) for each decoupled part so that each GD can be performed in a more specialized and effective manner. A GD-Unit consists of a dynamic graph where each vertice represents a modality and each edge indicates a dynamic knowledge distillation. Such GD paradigm provides a flexible knowledge transfer manner where the distillation weights can be automatically learned, thus enabling diverse crossmodal knowledge transfer patterns. Experimental results show DMD consistently obtains superior performance than state-of-the-art MER methods. Visualization results show the graph edges in DMD exhibit meaningful distributional patterns w.r.t. the modality-irrelevant/-exclusive feature spaces. Codes are released at https://github.com/mdswyz/DMD.

Temporal action detection aims to predict the time intervals and the classes of action instances in the video. Despite the promising performance, existing two-stream models exhibit slow inference speed due to their reliance on computationally expensive optical flow. In this paper, we introduce a decomposed cross-modal distillation framework to build a strong RGB-based detector by transferring knowledge of the motion modality. Specifically, instead of direct distillation, we propose to separately learn RGB and motion representations, which are in turn combined to perform action localization. The dual-branch design and the asymmetric training objectives enable effective motion knowledge transfer while preserving RGB information intact. In addition, we introduce a local attentive fusion to better exploit the multimodal complementarity. It is designed to preserve the local discriminability of the features that is important for action localization. Extensive experiments on the benchmarks verify the effectiveness of the proposed method in enhancing RGB-based action detectors. Notably, our framework is agnostic to backbones and detection heads, bringing consistent gains across different model combinations.

Encouraged by the effectiveness of encoding temporal dynamics within the frequency domain, recent human motion prediction systems prefer to first convert the motion representation from the original pose space into the frequency space. In this paper, we introduce two closer looks at effective frequency representation learning for robust motion prediction and summarize them as: decompose more and aggregate better. Motivated by these two insights, we develop two powerful units that factorize the frequency representation learning task with a novel decomposition-aggregation two-stage strategy: (1) frequency decomposition unit unweaves multi-view frequency representations from an input body motion by embedding its frequency features into multiple spaces; (2) feature aggregation unit deploys a series of intra-space and inter-space feature aggregation layers to collect comprehensive frequency representations from these spaces for robust human motion prediction. As evaluated on large-scale datasets, we develop a strong baseline model for the human motion prediction task that outperforms state-of-the-art methods by large margins: 8% 12% on Human3.6M, 3% 7% on CMU MoCap, and 7% 10% on 3DPW.

Understanding dense action in videos is a fundamental challenge towards the generalization of vision models. Several works show that compositionality is key to achieving generalization by combining known primitive elements, especially for handling novel composited structures. Compositional temporal grounding is the task of localizing dense action by using known words combined in novel ways in the form of novel query sentences for the actual grounding. In recent works, composition is assumed to be learned from pairs of whole videos and language embeddings through large scale self-supervised pre-training. Alternatively, one can process the video and language into word-level primitive elements, and then only learn fine-grained semantic correspondences. Both approaches do not consider the granularity of the compositions, where different query granularity corresponds to different video segments. Therefore, a good compositional representation should be sensitive to different video and query granularity. We propose a method to learn a coarse-to-fine compositional representation by decomposing the original query sentence into different granular levels, and then learning the correct correspondences between the video and recombined queries through a contrastive ranking constraint. Additionally, we run temporal boundary prediction in a coarse-to-fine manner for precise grounding boundary detection. Experiments are performed on two datasets Charades-CG and ActivityNet-CG showing the superior compositional generalizability of our approach.

Large pre-trained vision-language models (VLMs) reduce the time for developing predictive models for various vision-grounded language downstream tasks by providing rich, adaptable image and text representations. However, these models suffer from societal biases owing to the skewed distribution of various identity groups in the training data. These biases manifest as the skewed similarity between the representations for specific text concepts and images of people of different identity groups and, therefore, limit the usefulness of such models in real-world high-stakes applications. In this work, we present DeAR (Debiasing with Additive Residuals), a novel debiasing method that learns additive residual image representations to offset the original representations, ensuring fair output representations. In doing so, it reduces the ability of the representations to distinguish between the different identity groups. Further, we observe that the current fairness tests are performed on limited face image datasets that fail to indicate why a specific text concept should/should not apply to them. To bridge this gap and better evaluate DeAR, we introduce a new context-based bias benchmarking dataset - the Protected Attribute Tag Association (PATA) dataset for evaluating the fairness of large pre-trained VLMs. Additionally, PATA provides visual context for a diverse human population in different scenarios with both positive and negative connotations. Experimental results for fairness and zero-shot performance preservation using multiple datasets demonstrate the efficacy of our framework.

One-shot video-driven talking face generation aims at producing a synthetic talking video by transferring the facial motion from a video to an arbitrary portrait image. Head pose and facial expression are always entangled in facial motion and transferred simultaneously. However, the entanglement sets up a barrier for these methods to be used in video portrait editing directly, where it may require to modify the expression only while maintaining the pose unchanged. One challenge of decoupling pose and expression is the lack of paired data, such as the same pose but different expressions. Only a few methods attempt to tackle this challenge with the feat of 3D Morphable Models (3DMMs) for explicit disentanglement. But 3DMMs are not accurate enough to capture facial details due to the limited number of Blendshapes, which has side effects on motion transfer. In this paper, we introduce a novel self-supervised disentanglement framework to decouple pose and expression without 3DMMs and paired data, which consists of a motion editing module, a pose generator, and an expression generator. The editing module projects faces into a latent space where pose motion and expression motion can be disentangled, and the pose or expression transfer can be performed in the latent space conveniently via addition. The two generators render the modified latent codes to images, respectively. Moreover, to guarantee the disentanglement, we propose a bidirectional cyclic training strategy with well-designed constraints. Evaluations demonstrate our method can control pose or expression independently and be used for general video editing.

The exclusive properties of RAW data have shown great potential for low-light image enhancement. Nevertheless, the performance is bottlenecked by the inherent limitations of existing architectures in both single-stage and multi-stage methods. Mixed mapping across two different domains, noise-to-clean and RAW-to-sRGB, misleads the single-stage methods due to the domain ambiguity. The multi-stage methods propagate the information merely through the resulting image of each stage, neglecting the abundant features in the lossy image-level dataflow. In this paper, we probe a generalized solution to these bottlenecks and propose a Decouple aNd Feedback framework, abbreviated as DNF. To mitigate the domain ambiguity, domainspecific subtasks are decoupled, along with fully utilizing the unique properties in RAW and sRGB domains. The feature propagation across stages with a feedback mechanism avoids the information loss caused by image-level dataflow. The two key insights of our method resolve the inherent limitations of RAW data-based low-light image enhancement satisfactorily, empowering our method to outperform the previous state-of-the-art method by a large margin with only 19% parameters, achieving 0.97dB and 1.30dB PSNR improvements on the Sony and Fuji subsets of SID.

With the rapid development of virtual reality, 360deg images have gained increasing popularity. Their wide field of view necessitates high resolution to ensure image quality. This, however, makes it harder to acquire, store and even process such 360deg images. To alleviate this issue, we propose the first attempt at 360deg image rescaling, which refers to downscaling a 360deg image to a visually valid low-resolution (LR) counterpart and then upscaling to a high-resolution (HR) 360deg image given the LR variant. Specifically, we first analyze two 360deg image datasets and observe several findings that characterize how 360deg images typically change along their latitudes. Inspired by these findings, we propose a novel deformable invertible neural network (INN), named DINN360, for latitude-aware 360deg image rescaling. In DINN360, a deformable INN is designed to downscale the LR image, and project the high-frequency (HF) component to the latent space by adaptively handling various deformations occurring at different latitude regions. Given the downscaled LR image, the high-quality HR image is then reconstructed in a conditional latitude-aware manner by recovering the structure-related HF component from the latent space. Extensive experiments over four public datasets show that our DINN360 method performs considerably better than other state-of-the-art methods for 2x, 4x and 8x 360deg image rescaling.

Implicit neural representation (INR) characterizes the attributes of a signal as a function of corresponding coordinates which emerges as a sharp weapon for solving inverse problems. However, the capacity of INR is limited by the spectral bias in the network training. In this paper, we find that such a frequency-related problem could be largely solved by re-arranging the coordinates of the input signal, for which we propose the disorder-invariant implicit neural representation (DINER) by augmenting a hash-table to a traditional INR backbone. Given discrete signals sharing the same histogram of attributes and different arrangement orders, the hash-table could project the coordinates into the same distribution for which the mapped signal can be better modeled using the subsequent INR network, leading to significantly alleviated spectral bias. Experiments not only reveal the generalization of the DINER for different INR backbones (MLP vs. SIREN) and various tasks (image/video representation, phase retrieval, and refractive index recovery) but also show the superiority over the state-of-the-art algorithms both in quality and speed.

Considering the ill-posed nature, contrastive regularization has been developed for single image dehazing, introducing the information from negative images as a lower bound. However, the contrastive samples are nonconsensual, as the negatives are usually represented distantly from the clear (i.e., positive) image, leaving the solution space still under-constricted. Moreover, the interpretability of deep dehazing models is underexplored towards the physics of the hazing process. In this paper, we propose a novel curricular contrastive regularization targeted at a consensual contrastive space as opposed to a non-consensual one. Our negatives, which provide better lower-bound constraints, can be assembled from 1) the hazy image, and 2) corresponding restorations by other existing methods. Further, due to the different similarities between the embeddings of the clear image and negatives, the learning difficulty of the multiple components is intrinsically imbalanced. To tackle this issue, we customize a curriculum learning strategy to reweight the importance of different negatives. In addition, to improve the interpretability in the feature space, we build a physics-aware dual-branch unit according to the atmospheric scattering model. With the unit, as well as curricular contrastive regularization, we establish our dehazing network, named C2PNet. Extensive experiments demonstrate that our C2PNet significantly outperforms state-of-the-art methods, with extreme PSNR boosts of 3.94dB and 1.50dB, respectively, on SOTS-indoor and SOTS-outdoor datasets. Code is available at https://github.com/YuZheng9/C2PNet.

Supervised crowd counting relies heavily on costly manual labeling, which is difficult and expensive, especially in dense scenes. To alleviate the problem, we propose a novel unsupervised framework for crowd counting, named CrowdCLIP. The core idea is built on two observations: 1) the recent contrastive pre-trained vision-language model (CLIP) has presented impressive performance on various downstream tasks; 2) there is a natural mapping between crowd patches and count text. To the best of our knowledge, CrowdCLIP is the first to investigate the vision-language knowledge to solve the counting problem. Specifically, in the training stage, we exploit the multi-modal ranking loss by constructing ranking text prompts to match the size-sorted crowd patches to guide the image encoder learning. In the testing stage, to deal with the diversity of image patches, we propose a simple yet effective progressive filtering strategy to first select the highly potential crowd patches and then map them into the language space with various counting intervals. Extensive experiments on five challenging datasets demonstrate that the proposed CrowdCLIP achieves superior performance compared to previous unsupervised state-of-the-art counting methods. Notably, CrowdCLIP even surpasses some popular fully-supervised methods under the cross-dataset setting. The source code will be available at https://github.com/dk-liang/CrowdCLIP.

Image-based multi-person reconstruction in wide-field large scenes is critical for crowd analysis and security alert. However, existing methods cannot deal with large scenes containing hundreds of people, which encounter the challenges of large number of people, large variations in human scale, and complex spatial distribution. In this paper, we propose Crowd3D, the first framework to reconstruct the 3D poses, shapes and locations of hundreds of people with global consistency from a single large-scene image. The core of our approach is to convert the problem of complex crowd localization into pixel localization with the help of our newly defined concept, Human-scene Virtual Interaction Point (HVIP). To reconstruct the crowd with global consistency, we propose a progressive reconstruction network based on HVIP by pre-estimating a scene-level camera and a ground plane. To deal with a large number of persons and various human sizes, we also design an adaptive human-centric cropping scheme. Besides, we contribute a benchmark dataset, LargeCrowd, for crowd reconstruction in a large scene. Experimental results demonstrate the effectiveness of the proposed method. The code and the dataset are available at http://cic.tju.edu.cn/faculty/likun/projects/Crowd3D.