CVPR 2022 Papers

Recently, vision Transformers (ViTs) are developing rapidly and starting to challenge the domination of convolutional neural networks (CNNs) in the realm of computer vision (CV). With the general-purpose Transformer architecture replacing the hard-coded inductive biases of convolution, ViTs have surpassed CNNs, especially in data-sufficient circumstances. However, ViTs are prone to over-fit on small datasets and thus rely on large-scale pre-training, which expends enormous time. In this paper, we strive to liberate ViTs from pre-training by introducing CNNs' inductive biases back to ViTs while preserving their network architectures for higher upper bound and setting up more suitable optimization objectives. To begin with, an agent CNN is designed based on the given ViT with inductive biases. Then a bootstrapping training algorithm is proposed to jointly optimize the agent and ViT with weight sharing, during which the ViT learns inductive biases from the intermediate features of the agent. Extensive experiments on CIFAR-10/100 and ImageNet-1k with limited training data have shown encouraging results that the inductive biases help ViTs converge significantly faster and outperform conventional CNNs with even fewer parameters. Our code is publicly available at https://github.com/zhfeing/Bootstrapping-ViTs-pytorch.

Volumetric view synthesis methods with neural representations, such as NeRF and NeX, have recently demonstrated high-quality novel view synthesis. Optimizing these representations is slow, however, and even fully trained models cannot reproduce all fine details in the input views. We present a simple but effective technique to boost the rendering quality, which can be easily integrated with most view synthesis methods. The core idea is to adaptively transfer color residuals (the difference between the input images and their reconstruction) from training views to novel views. We blend the residuals from multiple views using a heuristic weighting scheme depending on ray visibility and angular differences. We integrate our technique with several state-of-the-art view synthesis methods and evaluate on the Real Forward-facing and the Shiny datasets. Our results show that at about 1/10th the number of training iterations we achieve the same rendering quality as fully converged NeRF and NeX models, and when applied to fully converged models we significantly improve their rendering quality.

Deep image matting methods have achieved increasingly better results on benchmarks (e.g., Composition-1k/alphamatting.com). However, the robustness, including robustness to trimaps and generalization to images from different domains, is still under-explored. Although some works propose to either refine the trimaps or adapt the algorithms to real-world images via extra data augmentation, none of them has taken both into consideration, not to mention the significant performance deterioration on benchmarks while using those data augmentation. To fill this gap, we propose an image matting method which achieves higher robustness (RMat) via multilevel context assembling and strong data augmentation targeting matting. Specifically, we first build a strong matting framework by modeling ample global information with transformer blocks in the encoder, and focusing on details in combination with convolution layers as well as a low-level feature assembling attention block in the decoder. Then, based on this strong baseline, we analyze current data augmentation and explore simple but effective strong data augmentation to boost the baseline model and contribute a more generalizable matting method. Compared with previous methods, the proposed method not only achieves state-of-the-art results on the Composition-1k benchmark (11% improvement on SAD and 27% improvement on Grad) with smaller model size, but also shows more robust generalization results on other benchmarks, on real-world images, and also on varying coarse-to-fine trimaps with our extensive experiments.

This paper focuses on crowd counting. As large-scale variations often exist within crowd images, neither fixed-size convolution kernel of CNN nor fixed-size attentions of recent vision transformers can well handle this kind of variations. To address this problem, we propose a Multifaceted Attention Network (MAN), which incorporates global attention from vanilla transformer, learnable local attention, attention regularization and instance attention into a counting model. Firstly, the local Learnable Region Attention (LRA) is proposed to assign attention exclusive for each feature location dynamically. Secondly, we design the Local Attention Regularization to supervise the training of LRA by minimizing the deviation among the attention for different feature locations. Finally, we provide an Instance Attention mechanism to focus on the most important instances dynamically during training. Extensive experiments on four challenging crowd counting datasets namely ShanghaiTech, UCF-QNRF, JHU++, and NWPU have validated the proposed method.

This work studies black-box adversarial attacks against deep neural networks (DNNs), where the attacker can only access the query feedback returned by the attacked DNN model, while other information such as model parameters or the training datasets are unknown. One promising approach to improve attack performance is utilizing the adversarial transferability between some white-box surrogate models and the target model (i.e., the attacked model). However, due to the possible differences on model architectures and training datasets between surrogate and target models, dubbed "surrogate biases", the contribution of adversarial transferability to improving the attack performance may be weakened. To tackle this issue, we innovatively propose a black-box attack method by developing a novel mechanism of adversarial transferability, which is robust to the surrogate biases. The general idea is transferring partial parameters of the conditional adversarial distribution (CAD) of surrogate models, while learning the untransferred parameters based on queries to the target model, to keep the flexibility to adjust the CAD of the target model on any new benign sample. Extensive experiments on benchmark datasets and attacking against real-world API demonstrate the superior attack performance of the proposed method.

This paper presents a new approach to boost a single-modality (LiDAR) 3D object detector by teaching it to simulate features and responses that follow a multi-modality (LiDAR-image) detector. The approach needs LiDAR-image data only when training the single-modality detector, and once well-trained, it only needs LiDAR data at inference. We design a novel framework to realize the approach: response distillation to focus on the crucial response samples and avoid the background samples; sparse-voxel distillation to learn voxel semantics and relations from the estimated crucial voxels; a fine-grained voxel-to-point distillation to better attend to features of small and distant objects; and instance distillation to further enhance the deep-feature consistency. Experimental results on the nuScenes dataset show that our approach outperforms all SOTA LiDAR-only 3D detectors and even surpasses the baseline LiDAR-image detector on the key NDS metric, filling 72% mAP gap between the single- and multi-modality detectors.

Deep learning (DL) models trained to minimize empirical risk on a single domain often fail to generalize when applied to other domains. Model failures due to poor generalizability are quite common in practice and may prove quite perilous in mission-critical applications, e.g., diagnostic imaging where real-world data often exhibits pronounced variability. Such limitations have led to increased interest in domain generalization (DG) approaches that improve the ability of models learned from a single or multiple source domains to generalize to out-of-distribution (OOD) test domains. In this work, we propose BoosterNet, a lean add-on network that can be simply appended to any arbitrary core network to improve its generalization capability without requiring any changes in its architecture or training procedure. Specifically, using a novel measure of feature culpability, BoosterNet is trained episodically on the most and least culpable data features extracted from critical units in the core network based on their contribution towards class-specific prediction errors, which have shown to improve generalization. At inference time, corresponding test image features are extracted from the closest class-specific units, determined by smart gating via a Siamese network, and fed to BoosterNet for improved generalization. We evaluate the performance of BoosterNet within two very different classification problems, digits and skin lesions, and demonstrate a marked improvement in model generalization to OOD test domains compared to SOTA.

In this paper, we propose a novel semi-supervised learning (SSL) framework named BoostMIS that combines adaptive pseudo labeling and informative active annotation to unleash the potential of medical image SSL models: (1) BoostMIS can adaptively leverage the cluster assumption and consistency regularization of the unlabeled data according to the current learning status. This strategy can adaptively generate one-hot "hard" labels converted from task model predictions for better task model training. (2) For the unselected unlabeled images with low confidence, we introduce an Active learning (AL) algorithm to find the informative samples as the annotation candidates by exploiting virtual adversarial perturbation and model's density-aware entropy. These informative candidates are subsequently fed into the next training cycle for better SSL label propagation. Notably, the adaptive pseudo-labeling and informative active annotation form a learning closed-loop that are mutually collaborative to boost medical image SSL. To verify the effectiveness of the proposed method, we collected a metastatic epidural spinal cord compression (MESCC) dataset that aims to optimize MESCC diagnosis and classification for improved specialist referral and treatment. We conducted an extensive experimental study of BoostMIS on MESCC and another public dataset COVIDx. The experimental results verify our framework's effectiveness and generalisability for different medical image datasets with a significant improvement over various state-of-the-art methods.

A significant gap remains between today's visual pattern recognition models and human-level visual cognition especially when it comes to few-shot learning and compositional reasoning of novel concepts. We introduce Bongard-HOI, a new visual reasoning benchmark that focuses on compositional learning of human-object interactions (HOIs) from natural images. It is inspired by two desirable characteristics from the classical Bongard problems (BPs): 1) few-shot concept learning, and 2) context-dependent reasoning. We carefully curate the few-shot instances with hard negatives, where positive and negative images only disagree on action labels, making mere recognition of object categories insufficient to complete our benchmarks. We also design multiple test sets to systematically study the generalization of visual learning models, where we vary the overlap of the HOI concepts between the training and test sets of few- shot instances, from partial to no overlaps. Bongard-HOI presents a substantial challenge to today's visual recognition models. The state-of-the-art HOI detection model achieves only 62% accuracy on few-shot binary prediction while even amateur human testers on MTurk have 91% accuracy. With the Bongard-HOI benchmark, we hope to further advance research efforts in visual reasoning, especially in holistic perception-reasoning systems and better representation learning.

We propose BokehMe, a hybrid bokeh rendering framework that marries a neural renderer with a classical physically motivated renderer. Given a single image and a potentially imperfect disparity map, BokehMe generates high-resolution photo-realistic bokeh effects with adjustable blur size, focal plane, and aperture shape. To this end, we analyze the errors from the classical scattering-based method and derive a formulation to calculate an error map. Based on this formulation, we implement the classical renderer by a scattering-based method and propose a two-stage neural renderer to fix the erroneous areas from the classical renderer. The neural renderer employs a dynamic multi-scale scheme to efficiently handle arbitrary blur sizes, and it is trained to handle imperfect disparity input. Experiments show that our method compares favorably against previous methods on both synthetic image data and real image data with predicted disparity. A user study is further conducted to validate the advantage of our method.

Dense correspondence between humans carries powerful semantic information that can be utilized to solve fundamental problems for full-body understanding such as in-the-wild surface matching, tracking and reconstruction. In this paper we present BodyMap, a new framework for obtaining high-definition full-body and continuous dense correspondence between in-the-wild images of clothed humans and the surface of a 3D template model. The correspondences cover fine details such as hands and hair, while capturing regions far from the body surface, such as loose clothing. Prior methods for estimating such dense surface correspondence i) cut a 3D body into parts which are unwrapped to a 2D UV space, producing discontinuities along part seams, or ii) use a single surface for representing the whole body, but none handled body details. Here, we introduce a novel network architecture with Vision Transformers that learn fine-level features on a continuous body surface. BodyMap outperforms prior work on various metrics and datasets, including DensePose-COCO by a large margin. Furthermore, we show various applications ranging from multi-layer dense cloth correspondence, neural rendering with novel-view synthesis and appearance swapping.

Recent advances in generative adversarial networks (GANs) have provided potential solutions for photorealistic human image synthesis. However, the explicit and individual control of synthesis over multiple factors, such as poses, body shapes, and skin colors, remains difficult for existing methods. This is because current methods mainly rely on a single pose/appearance model, which is limited in disentangling various poses and appearance in human images. In addition, such a unimodal strategy is prone to causing severe artifacts in the generated images like color distortions and unrealistic textures. To tackle these issues, this paper proposes a multi-factor conditioned method dubbed BodyGAN. Specifically, given a source image, our Body-GAN aims at capturing the characteristics of the human body from multiple aspects: (i) A pose encoding branch consisting of three hybrid subnetworks is adopted, to generate the semantic segmentation based representation, the 3D surface based representation, and the key point based representation of the human body, respectively. (ii) Based on the segmentation results, an appearance encoding branch is used to obtain the appearance information of the human body parts. (iii) The outputs of these two branches are represented by user-editable condition maps, which are then processed by a generator to predict the synthesized image. In this way, our BodyGAN can achieve the fine-grained disentanglement of pose, body shape, and appearance, and consequently enable the explicit and effective control of synthesis with diverse conditions. Extensive experiments on multiple datasets and a comprehensive user-study show that our BodyGAN achieves the state-of-the-art performance.

We present Block-NeRF, a variant of Neural Radiance Fields that can represent large-scale environments. Specifically, we demonstrate that when scaling NeRF to render city-scale scenes spanning multiple blocks, it is vital to decompose the scene into individually trained NeRFs. This decomposition decouples rendering time from scene size, enables rendering to scale to arbitrarily large environments, and allows per-block updates of the environment. We adopt several architectural changes to make NeRF robust to data captured over months under different environmental conditions. We add appearance embeddings, learned pose refinement, and controllable exposure to each individual NeRF, and introduce a procedure for aligning appearance between adjacent NeRFs so that they can be seamlessly combined. We build a grid of Block-NeRFs from 2.8 million images to create the largest neural scene representation to date, capable of rendering an entire neighborhood of San Francisco.

Real noisy-clean pairs on a large scale are costly and difficult to obtain. Meanwhile, supervised denoisers trained on synthetic data perform poorly in practice. Self-supervised denoisers, which learn only from single noisy images, solve the data collection problem. However, self-supervised denoising methods, especially blindspot-driven ones, suffer sizable information loss during input or network design. The absence of valuable information dramatically reduces the upper bound of denoising performance. In this paper, we propose a simple yet efficient approach called Blind2Unblind to overcome the information loss in blindspot-driven denoising methods. First, we introduce a global-aware mask mapper that enables global perception and accelerates training. The mask mapper samples all pixels at blind spots on denoised volumes and maps them to the same channel, allowing the loss function to optimize all blind spots at once. Second, we propose a re-visible loss to train the denoising network and make blind spots visible. The denoiser can learn directly from raw noise images without losing information or being trapped in identity mapping. We also theoretically analyze the convergence of the re-visible loss. Extensive experiments on synthetic and real-world datasets demonstrate the superior performance of our approach compared to previous work. Code is available at https://github.com/demonsjin/Blind2Unblind.

While researches on model-based blind single image super-resolution (SISR) have achieved tremendous successes recently, most of them do not consider the image degradation sufficiently. Firstly, they always assume image noise obeys an independent and identically distributed (i.i.d.) Gaussian or Laplacian distribution, which largely underestimates the complexity of real noise. Secondly, previous commonly-used kernel priors (e.g., normalization, sparsity) are not effective enough to guarantee a rational kernel solution, and thus degenerates the performance of subsequent SISR task. To address the above issues, this paper proposes a model-based blind SISR method under the probabilistic framework, which elaborately models image degradation from the perspectives of noise and blur kernel. Specifically, instead of the traditional i.i.d. noise assumption, a patch-based non-i.i.d. noise model is proposed to tackle the complicated real noise, expecting to increase the degrees of freedom of the model for noise representation. As for the blur kernel, we novelly con- struct a concise yet effective kernel generator, and plug it into the proposed blind SISR method as an explicit kernel prior (EKP). To solve the proposed model, a theoretically grounded Monte Carlo EM algorithm is specifically designed. Comprehensive experiments demonstrate the superiority of our method over current state-of-the-arts on synthetic and real datasets. The source code is available at https://github.com/zsyOAOA/BSRDM.

Blind face restoration, which aims to reconstruct high-quality images from low-quality inputs, can benefit many applications. Although existing generative-based methods achieve significant progress in producing high-quality images, they often fail to restore natural face shapes and high-fidelity facial details from severely-degraded inputs. In this work, we propose to integrate shape and generative priors to guide the challenging blind face restoration. Firstly, we set up a shape restoration module to recover reasonable facial geometry with 3D reconstruction. Secondly, a pretrained facial generator is adopted as decoder to generate photo-realistic high-resolution images. To ensure high-fidelity, hierarchical spatial features extracted from the low-quality inputs and rendered 3D images are inserted into the decoder with our proposed Adaptive Feature Fusion Block (AFFB). Moreover, we introduce hybrid-level losses to jointly train the shape and generative priors together with other network parts such that these two priors better adapt to our blind face restoration task. The proposed Shape and Generative Prior integrated Network (SGPN) can restore high-quality images with clear face shapes and realistic facial details. Experimental results on synthetic and real-world datasets demonstrate SGPN performs favorably against state-of-the-art blind face restoration methods.

Natural language offers a highly intuitive interface for image editing. In this paper, we introduce the first solution for performing local (region-based) edits in generic natural images, based on a natural language description along with an ROI mask. We achieve our goal by leveraging and combining a pretrained language-image model (CLIP), to steer the edit towards a user-provided text prompt, with a denoising diffusion probabilistic model (DDPM) to generate natural-looking results. To seamlessly fuse the edited region with the unchanged parts of the image, we spatially blend noised versions of the input image with the local text-guided diffusion latent at a progression of noise levels. In addition, we show that adding augmentations to the diffusion process mitigates adversarial results. We compare against several baselines and related methods, both qualitatively and quantitatively, and show that our method outperforms these solutions in terms of overall realism, ability to preserve the background and matching the text. Finally, we show several text-driven editing applications, including adding a new object to an image, removing/replacing/altering existing objects, background replacement, and image extrapolation.

Eulerian video magnification (EVM) has progressed to magnify subtle motions with a target frequency even under the presence of large motions of objects. However, existing EVM methods often fail to produce desirable results in real videos due to (1) mis-extracting subtle motions with a non-target frequency and (2) collapsing results when large de/acceleration motions occur (e.g., objects suddenly start, stop, or change direction). To enhance EVM performance on real videos, this paper proposes a bilateral video magnification filter (BVMF) that offers simple yet robust temporal filtering. BVMF has two kernels; (I) one kernel performs temporal bandpass filtering via a Laplacian of Gaussian whose passband peaks at the target frequency with unity gain and (II) the other kernel excludes large motions outside the magnitude of interest by Gaussian filtering on the intensity of the input signal via the Fourier shift theorem. Thus, BVMF extracts only subtle motions with the target frequency while excluding large motions outside the magnitude of interest, regardless of motion dynamics. In addition, BVMF runs the two kernels in the temporal and intensity domains simultaneously like the bilateral filter does in the spatial and intensity domains. This simplifies implementation and, as a secondary effect, keeps the memory usage low. Experiments conducted on synthetic and real videos show that BVMF outperforms state-of-the-art methods.

Shadow removal, which aims to restore the background in the shadow regions, is challenging due to the highly ill-posed nature. Most existing deep learning-based methods individually remove the shadow by only considering the content of the matched paired images, barely taking into account the auxiliary supervision of shadow generation on shadow removal procedure. In this work, we argue that shadow removal and generation are interrelated and could provide useful informative supervision for each other. Specifically, we propose a new Bijective Mapping Network (BMNet), which couples the learning procedures of shadow removal and shadow generation in a unified parameter-shared framework. With consistent two-way constraints and synchronous optimization of the two procedures, BMNet could effectively recover the underlying background contents during the forward shadow removal procedure. In addition, through statistic analysis on real-world datasets, we observe and verify that shadow appearances under different color spectrums are inconsistent. This motivates us to design a Shadow-Invariant Color Guidance Module (SICGM), which can explicitly utilize the learned shadow-invariant color information to guide network color restoration, thereby further reducing color-bias effects. Experiments on the representative ISTD, ISTD+ and SRD benchmarks show that our proposed network outperforms the state-of-the-art method [??] in de-shadowing performance, while only using its 0.25% network parameters and 6.25% floating point operations (FLOPs).