CVPR 2023 Papers

Large-scale vision-language models (VLMs) pre-trained on billion-level data have learned general visual representations and broad visual concepts. In principle, the well-learned knowledge structure of the VLMs should be inherited appropriately when being transferred to downstream tasks with limited data. However, most existing efficient transfer learning (ETL) approaches for VLMs either damage or are excessively biased towards the prior knowledge, e.g., prompt tuning (PT) discards the pre-trained text-based classifier and builds a new one while adapter-style tuning (AT) fully relies on the pre-trained features. To address this, we propose a new efficient tuning approach for VLMs named Task Residual Tuning (TaskRes), which performs directly on the text-based classifier and explicitly decouples the prior knowledge of the pre-trained models and new knowledge regarding a target task. Specifically, TaskRes keeps the original classifier weights from the VLMs frozen and obtains a new classifier for the target task by tuning a set of prior-independent parameters as a residual to the original one, which enables reliable prior knowledge preservation and flexible task-specific knowledge exploration. The proposed TaskRes is simple yet effective, which significantly outperforms previous ETL methods (e.g., PT and AT) on 11 benchmark datasets while requiring minimal effort for the implementation. Our code is available at https://github.com/geekyutao/TaskRes.

The general domain of video segmentation is currently fragmented into different tasks spanning multiple benchmarks. Despite rapid progress in the state-of-the-art, current methods are overwhelmingly task-specific and cannot conceptually generalize to other tasks. Inspired by recent approaches with multi-task capability, we propose TarViS: a novel, unified network architecture that can be applied to any task that requires segmenting a set of arbitrarily defined 'targets' in video. Our approach is flexible with respect to how tasks define these targets, since it models the latter as abstract 'queries' which are then used to predict pixel-precise target masks. A single TarViS model can be trained jointly on a collection of datasets spanning different tasks, and can hot-swap between tasks during inference without any task-specific retraining. To demonstrate its effectiveness, we apply TarViS to four different tasks, namely Video Instance Segmentation (VIS), Video Panoptic Segmentation (VPS), Video Object Segmentation (VOS) and Point Exemplar-guided Tracking (PET). Our unified, jointly trained model achieves state-of-the-art performance on 5/7 benchmarks spanning these four tasks, and competitive performance on the remaining two. Code and model weights are available at: https://github.com/Ali2500/TarViS

We present a novel image inversion framework and a training pipeline to achieve high-fidelity image inversion with high-quality attribute editing. Inverting real images into StyleGAN's latent space is an extensively studied problem, yet the trade-off between the image reconstruction fidelity and image editing quality remains an open challenge. The low-rate latent spaces are limited in their expressiveness power for high-fidelity reconstruction. On the other hand, high-rate latent spaces result in degradation in editing quality. In this work, to achieve high-fidelity inversion, we learn residual features in higher latent codes that lower latent codes were not able to encode. This enables preserving image details in reconstruction. To achieve high-quality editing, we learn how to transform the residual features for adapting to manipulations in latent codes. We train the framework to extract residual features and transform them via a novel architecture pipeline and cycle consistency losses. We run extensive experiments and compare our method with state-of-the-art inversion methods. Qualitative metrics and visual comparisons show significant improvements.

Recent researches reveal that StyleGAN can generate highly realistic images, inspiring researchers to use pretrained StyleGAN to generate high-fidelity swapped faces. However, existing methods fail to meet the expectations in two essential aspects of high-fidelity face swapping. Their results are blurry without pore-level details and fail to preserve identity for challenging cases. To overcome the above artifacts, we innovatively construct a series of identity-preserving semantic bases of StyleGAN (called StyleIPSB) in respect of pose, expression, and illumination. Each basis of StyleIPSB controls one specific semantic attribute and disentangles with the others. The StyleIPSB constrains style code in the subspace of W+ space to preserve pore-level details. StyleIPSB gives us a novel tool for high-fidelity face swapping, and we propose a three-stage framework for face swapping with StyleIPSB. Firstly, we transform the target facial images' attributes to the source image. We learn the mapping from 3D Morphable Model (3DMM) parameters, which capture the prominent semantic variance, to the coordinates of StyleIPSB that show higher identity-preserving and fidelity. Secondly, to transform detailed attributes which 3DMM does not capture, we learn the residual attribute between the reenacted face and the target face. Finally, the face is blended into the background of the target image. Extensive results and comparisons demonstrate that StyleIPSB can effectively preserve identity and pore-level details. The results of face swapping can achieve state-of-the-art performance. We will release our code at https://github.com/a686432/StyleIPSB.

The high computational costs of video super-resolution (VSR) models hinder their deployment on resource-limited devices, e.g., smartphones and drones. Existing VSR models contain considerable redundant filters, which drag down the inference efficiency. To prune these unimportant filters, we develop a structured pruning scheme called Structured Sparsity Learning (SSL) according to the properties of VSR. In SSL, we design pruning schemes for several key components in VSR models, including residual blocks, recurrent networks, and upsampling networks. Specifically, we develop a Residual Sparsity Connection (RSC) scheme for residual blocks of recurrent networks to liberate pruning restrictions and preserve the restoration information. For upsampling networks, we design a pixel-shuffle pruning scheme to guarantee the accuracy of feature channel-space conversion. In addition, we observe that pruning error would be amplified as the hidden states propagate along with recurrent networks. To alleviate the issue, we design Temporal Finetuning (TF). Extensive experiments show that SSL can significantly outperform recent methods quantitatively and qualitatively. The code is available at https://github.com/Zj-BinXia/SSL.

Images taken in a low light condition with the presence of camera shake suffer from motion blur and photon shot noise. While state-of-the-art image restoration networks show promising results, they are largely limited to well-illuminated scenes and their performance drops significantly when photon shot noise is strong. In this paper, we propose a new blur estimation technique customized for photon-limited conditions. The proposed method employs a gradient-based backpropagation method to estimate the blur kernel. By modeling the blur kernel using a low-dimensional representation with the key points on the motion trajectory, we significantly reduce the search space and improve the regularity of the kernel estimation problem. When plugged into an iterative framework, our novel low-dimensional representation provides improved kernel estimates and hence significantly better deconvolution performance when compared to end-to-end trained neural networks.

To obtain clean images with salient structures from noisy observations, a growing trend in current denoising studies is to seek the help of additional guidance images with high signal-to-noise ratios, which are often acquired in different spectral bands such as near infrared. Although previous guided denoising methods basically require the input images to be well-aligned, a more common way to capture the paired noisy target and guidance images is to exploit a stereo camera system. However, current studies on cross-spectral stereo matching cannot fully guarantee the pixel-level registration accuracy, and rarely consider the case of noise contamination. In this work, for the first time, we propose a guided denoising framework for cross-spectral stereo images. Instead of aligning the input images via conventional stereo matching, we aggregate structures from the guidance image to estimate a clean structure map for the noisy target image, which is then used to regress the final denoising result with a spatially variant linear representation model. Based on this, we design a neural network, called as SANet, to complete the entire guided denoising process. Experimental results show that, our SANet can effectively transfer structures from an unaligned guidance image to the restoration result, and outperforms state-of-the-art denoisers on various stereo image datasets. Besides, our structure aggregation strategy also shows its potential to handle other unaligned guided restoration tasks such as super-resolution and deblurring. The source code is available at https://github.com/lustrouselixir/SANet.

Video understanding tasks have traditionally been modeled by two separate architectures, specially tailored for two distinct tasks. Sequence-based video tasks, such as action recognition, use a video backbone to directly extract spatiotemporal features, while frame-based video tasks, such as multiple object tracking (MOT), rely on single fixed-image backbone to extract spatial features. In contrast, we propose to unify video understanding tasks into one novel streaming video architecture, referred to as Streaming Vision Transformer (S-ViT). S-ViT first produces frame-level features with a memory-enabled temporally-aware spatial encoder to serve the frame-based video tasks. Then the frame features are input into a task-related temporal decoder to obtain spatiotemporal features for sequence-based tasks. The efficiency and efficacy of S-ViT is demonstrated by the state-of-the-art accuracy in the sequence-based action recognition task and the competitive advantage over conventional architecture in the frame-based MOT task. We believe that the concept of streaming video model and the implementation of S-ViT are solid steps towards a unified deep learning architecture for video understanding. Code will be available at https://github.com/yuzhms/Streaming-Video-Model.

While recent methods for motion and stereo estimation recover an unprecedented amount of details, such highly detailed structures are neither adequately reflected in the data of existing benchmarks nor their evaluation methodology. Hence, we introduce Spring -- a large, high-resolution, high-detail, computer-generated benchmark for scene flow, optical flow, and stereo. Based on rendered scenes from the open-source Blender movie "Spring", it provides photo-realistic HD datasets with state-of-the-art visual effects and ground truth training data. Furthermore, we provide a website to upload, analyze and compare results. Using a novel evaluation methodology based on a super-resolved UHD ground truth, our Spring benchmark can assess the quality of fine structures and provides further detailed performance statistics on different image regions. Regarding the number of ground truth frames, Spring is 60x larger than the only scene flow benchmark, KITTI 2015, and 15x larger than the well-established MPI Sintel optical flow benchmark. Initial results for recent methods on our benchmark show that estimating fine details is indeed challenging, as their accuracy leaves significant room for improvement. The Spring benchmark and the corresponding datasets are available at http://spring-benchmark.org.

Denoising is a crucial step for hyperspectral image (HSI) applications. Though witnessing the great power of deep learning, existing HSI denoising methods suffer from limitations in capturing the non-local self-similarity. Transformers have shown potential in capturing long-range dependencies, but few attempts have been made with specifically designed Transformer to model the spatial and spectral correlation in HSIs. In this paper, we address these issues by proposing a spectral enhanced rectangle Transformer, driving it to explore the non-local spatial similarity and global spectral low-rank property of HSIs. For the former, we exploit the rectangle self-attention horizontally and vertically to capture the non-local similarity in the spatial domain. For the latter, we design a spectral enhancement module that is capable of extracting global underlying low-rank property of spatial-spectral cubes to suppress noise, while enabling the interactions among non-overlapping spatial rectangles. Extensive experiments have been conducted on both synthetic noisy HSIs and real noisy HSIs, showing the effectiveness of our proposed method in terms of both objective metric and subjective visual quality. The code is available at https://github.com/MyuLi/SERT.

Recently deep learning techniques have significantly advanced image super-resolution (SR). Due to the black-box nature, quantifying reconstruction uncertainty is crucial when employing these deep SR networks. Previous approaches for SR uncertainty estimation mostly focus on capturing pixel-wise uncertainty in the spatial domain. SR uncertainty in the frequency domain which is highly related to image SR is seldom explored. In this paper, we propose to quantify spectral Bayesian uncertainty in image SR. To achieve this, a Dual-Domain Learning (DDL) framework is first proposed. Combined with Bayesian approaches, the DDL model is able to estimate spectral uncertainty accurately, enabling a reliability assessment for high frequencies reasoning from the frequency domain perspective. Extensive experiments under non-ideal premises are conducted and demonstrate the effectiveness of the proposed spectral uncertainty. Furthermore, we propose a novel Spectral Uncertainty based Decoupled Frequency (SUDF) training scheme for perceptual SR. Experimental results show the proposed SUDF can evidently boost perceptual quality of SR results without sacrificing much pixel accuracy.

Unsupervised Domain Adaptation (UDA) of semantic segmentation transfers labeled source knowledge to an unlabeled target domain by relying on accessing both the source and target data. However, the access to source data is often restricted or infeasible in real-world scenarios. Under the source data restrictive circumstances, UDA is less practical. To address this, recent works have explored solutions under the Source-Free Domain Adaptation (SFDA) setup, which aims to adapt a source-trained model to the target domain without accessing source data. Still, existing SFDA approaches use only image-level information for adaptation, making them sub-optimal in video applications. This paper studies SFDA for Video Semantic Segmentation (VSS), where temporal information is leveraged to address video adaptation. Specifically, we propose Spatio-Temporal Pixel-Level (STPL) contrastive learning, a novel method that takes full advantage of spatio-temporal information to tackle the absence of source data better. STPL explicitly learns semantic correlations among pixels in the spatio-temporal space, providing strong self-supervision for adaptation to the unlabeled target domain. Extensive experiments show that STPL achieves state-of-the-art performance on VSS benchmarks compared to current UDA and SFDA approaches. Code is available at: https://github.com/shaoyuanlo/STPL

Dual-pixel photography is monocular RGB-D photography with an ultra-high resolution, enabling many applications in computational photography. However, there are still several challenges to fully utilizing dual-pixel photography. Unlike the conventional stereo pair, the dual pixel exhibits a bidirectional disparity that includes positive and negative values, depending on the focus plane depth in an image. Furthermore, capturing a wide range of dual-pixel disparity requires a shallow depth of field, resulting in a severely blurred image, degrading depth estimation performance. Recently, several data-driven approaches have been proposed to mitigate these two challenges. However, due to the lack of the ground-truth dataset of the dual-pixel disparity, existing data-driven methods estimate either inverse depth or blurriness map. In this work, we propose a self-supervised learning method that learns bidirectional disparity by utilizing the nature of anisotropic blur kernels in dual-pixel photography. We observe that the dual-pixel left/right images have reflective-symmetric anisotropic kernels, so their sum is equivalent to that of a conventional image. We take a self-supervised training approach with the novel kernel-split symmetry loss accounting for the phenomenon. Our method does not rely on a training dataset of dual-pixel disparity that does not exist yet. Our method can estimate a complete disparity map with respect to the focus-plane depth from a dual-pixel image, outperforming the baseline dual-pixel methods.

Significant progress has been made in self-supervised image denoising (SSID) in the recent few years. However, most methods focus on dealing with spatially independent noise, and they have little practicality on real-world sRGB images with spatially correlated noise. Although pixel-shuffle downsampling has been suggested for breaking the noise correlation, it breaks the original information of images, which limits the denoising performance. In this paper, we propose a novel perspective to solve this problem, i.e., seeking for spatially adaptive supervision for real-world sRGB image denoising. Specifically, we take into account the respective characteristics of flat and textured regions in noisy images, and construct supervisions for them separately. For flat areas, the supervision can be safely derived from non-adjacent pixels, which are much far from the current pixel for excluding the influence of the noise-correlated ones. And we extend the blind-spot network to a blind-neighborhood network (BNN) for providing supervision on flat areas. For textured regions, the supervision has to be closely related to the content of adjacent pixels. And we present a locally aware network (LAN) to meet the requirement, while LAN itself is selectively supervised with the output of BNN. Combining these two supervisions, a denoising network (e.g., U-Net) can be well-trained. Extensive experiments show that our method performs favorably against state-of-the-art SSID methods on real-world sRGB photographs. The code is available at https://github.com/nagejacob/SpatiallyAdaptiveSSID.

In low-level video analyses, effective representations are important to derive the correspondences between video frames. These representations have been learned in a self-supervised fashion from unlabeled images/videos, using carefully designed pretext tasks in some recent studies. However, the previous work concentrates on either spatial-discriminative features or temporal-repetitive features, with little attention to the synergy between spatial and temporal cues. To address this issue, we propose a novel spatial-then-temporal self-supervised learning method. Specifically, we firstly extract spatial features from unlabeled images via contrastive learning, and secondly enhance the features by exploiting the temporal cues in unlabeled videos via reconstructive learning. In the second step, we design a global correlation distillation loss to ensure the learning not to forget the spatial cues, and we design a local correlation distillation loss to combat the temporal discontinuity that harms the reconstruction. The proposed method outperforms the state-of-the-art self-supervised methods, as established by the experimental results on a series of correspondence-based video analysis tasks. Also, we performed ablation studies to verify the effectiveness of the two-step design as well as the distillation losses.

Face super-resolution (FSR) aims to reconstruct high-resolution (HR) face images from the low-resolution (LR) ones. With the advent of deep learning, the FSR technique has achieved significant breakthroughs. However, existing FSR methods either have a fixed receptive field or fail to maintain facial structure, limiting the FSR performance. To circumvent this problem, Fourier transform is introduced, which can capture global facial structure information and achieve image-size receptive field. Relying on the Fourier transform, we devise a spatial-frequency mutual network (SFMNet) for FSR, which is the first FSR method to explore the correlations between spatial and frequency domains as far as we know. To be specific, our SFMNet is a two-branch network equipped with a spatial branch and a frequency branch. Benefiting from the property of Fourier transform, the frequency branch can achieve image-size receptive field and capture global dependency while the spatial branch can extract local dependency. Considering that these dependencies are complementary and both favorable for FSR, we further develop a frequency-spatial interaction block (FSIB) which mutually amalgamates the complementary spatial and frequency information to enhance the capability of the model. Quantitative and qualitative experimental results show that the proposed method outperforms state-of-the-art FSR methods in recovering face images. The implementation and model will be released at https://github.com/wcy-cs/SFMNet.

Vision Transformers (ViT) have shown competitive advantages in terms of performance compared to convolutional neural networks (CNNs), though they often come with high computational costs. To this end, previous methods explore different attention patterns by limiting a fixed number of spatially nearby tokens to accelerate the ViT's multi-head self-attention (MHSA) operations. However, such structured attention patterns limit the token-to-token connections to their spatial relevance, which disregards learned semantic connections from a full attention mask. In this work, we propose an approach to learn instance-dependent attention patterns, by devising a lightweight connectivity predictor module that estimates the connectivity score of each pair of tokens. Intuitively, two tokens have high connectivity scores if the features are considered relevant either spatially or semantically. As each token only attends to a small number of other tokens, the binarized connectivity masks are often very sparse by nature and therefore provide the opportunity to reduce network FLOPs via sparse computations. Equipped with the learned unstructured attention pattern, sparse attention ViT (Sparsifiner) produces a superior Pareto frontier between FLOPs and top-1 accuracy on ImageNet compared to token sparsity. Our method reduces 48% 69% FLOPs of MHSA while the accuracy drop is within 0.4%. We also show that combining attention and token sparsity reduces ViT FLOPs by over 60%.

Modern neural networks are over-parameterized and thus rely on strong regularization such as data augmentation and weight decay to reduce overfitting and improve generalization. The dominant form of data augmentation applies invariant transforms, where the learning target of a sample is invariant to the transform applied to that sample. We draw inspiration from human visual classification studies and propose generalizing augmentation with invariant transforms to soft augmentation where the learning target softens non-linearly as a function of the degree of the transform applied to the sample: e.g., more aggressive image crop augmentations produce less confident learning targets. We demonstrate that soft targets allow for more aggressive data augmentation, offer more robust performance boosts, work with other augmentation policies, and interestingly, produce better calibrated models (since they are trained to be less confident on aggressively cropped/occluded examples). Combined with existing aggressive augmentation strategies, soft targets 1) double the top-1 accuracy boost across Cifar-10, Cifar-100, ImageNet-1K, and ImageNet-V2, 2) improve model occlusion performance by up to 4x, and 3) half the expected calibration error (ECE). Finally, we show that soft augmentation generalizes to self-supervised classification tasks.

In order to tackle video semantic segmentation task at a lower cost, e.g., only one frame annotated per video, lots of efforts have been devoted to investigate the utilization of those unlabeled frames by either assigning pseudo labels or performing feature enhancement. In this work, we propose a novel feature enhancement network to simultaneously model short- and long-term temporal correlation. Compared with existing work that only leverage short-term correspondence, the long-term temporal correlation obtained from distant frames can effectively expand the temporal perception field and provide richer contextual prior. More importantly, modeling adjacent and distant frames together can alleviate the risk of over-fitting, hence produce high-quality feature representation for the distant unlabeled frames in training set and unseen videos in testing set. To this end, we term our method SSLTM, short for Simultaneously Short- and Long-Term Temporal Modeling. In the setting of only one frame annotated per video, SSLTM significantly outperforms the state-of-the-art methods by 2% 3% mIoU on the challenging VSPW dataset. Furthermore, when working with a pseudo label based method such as MeanTeacher, our final model only exhibits 0.13% mIoU less than the ceiling performance (i.e., all frames are manually annotated).