CVPR 2022 Papers

Optimization of image transformation functions for the purpose of data augmentation has been intensively studied. In particular, adversarial data augmentation strategies, which search augmentation maximizing task loss, show significant improvement in the model generalization for many tasks. However, the existing methods require careful parameter tuning to avoid excessively strong deformations that take away image features critical for acquiring generalization. In this paper, we propose a data augmentation optimization method based on the adversarial strategy called TeachAugment, which can produce informative transformed images to the model without requiring careful tuning by leveraging a teacher model. Specifically, the augmentation is searched so that augmented images are adversarial for the target model and recognizable for the teacher model. We also propose data augmentation using neural networks, which simplifies the search space design and allows for updating of the data augmentation using the gradient method. We show that TeachAugment outperforms existing methods in experiments of image classification, semantic segmentation, and unsupervised representation learning tasks.

Recognizing discriminative details such as eyes and beaks is important for distinguishing fine-grained classes since they have similar overall appearances. In this regard, we introduce Task Discrepancy Maximization (TDM), a simple module for fine-grained few-shot classification. Our objective is to localize the class-wise discriminative regions by highlighting channels encoding distinct information of the class. Specifically, TDM learns task-specific channel weights based on two novel components: Support Attention Module (SAM) and Query Attention Module (QAM). SAM produces a support weight to represent channel-wise discriminative power for each class. Still, since the SAM is basically only based on the labeled support sets, it can be vulnerable to bias toward such support set. Therefore, we propose QAM which complements SAM by yielding a query weight that grants more weight to object-relevant channels for a given query image. By combining these two weights, a class-wise task-specific channel weight is defined. The weights are then applied to produce task-adaptive feature maps more focusing on the discriminative details. Our experiments validate the effectiveness of TDM and its complementary benefits with prior methods in fine-grained few-shot classification.

Domain adaptive object detection (DAOD) is a promising way to alleviate performance drop of detectors in new scenes. Albeit great effort made in single source domain adaptation, a more generalized task with multiple source domains remains not being well explored, due to knowledge degradation during their combination. To address this issue, we propose a novel approach, namely target-relevant knowledge preservation (TRKP), to unsupervised multi-source DAOD. Specifically, TRKP adopts the teacher-student framework, where the multi-head teacher network is built to extract knowledge from labeled source domains and guide the student network to learn detectors in unlabeled target domain. The teacher network is further equipped with an adversarial multi-source disentanglement (AMSD) module to preserve source domain-specific knowledge and simultaneously perform cross-domain alignment. Besides, a holistic target-relevant mining (HTRM) scheme is developed to re-weight the source images according to the source-target relevance. By this means, the teacher network is enforced to capture target-relevant knowledge, thus benefiting decreasing domain shift when mentoring object detection in the target domain. Extensive experiments are conducted on various widely used benchmarks with new state-of-the-art scores reported, highlighting the effectiveness.

This study addresses the issue of fusing infrared and visible images that appear differently for object detection. Aiming at generating an image of high visual quality, previous approaches discover commons underlying the two modalities and fuse upon the common space either by iterative optimization or deep networks. These approaches neglect that modality differences implying the complementary information are extremely important for both fusion and subsequent detection task. This paper proposes a bilevel optimization formulation for the joint problem of fusion and detection, and then unrolls to a target-aware Dual Adversarial Learning (TarDAL) network for fusion and a commonly used detection network. The fusion network with one generator and dual discriminators seeks commons while learning from differences, which preserves structural information of targets from the infrared and textural details from the visible. Furthermore, we build a synchronized imaging system with calibrated infrared and optical sensors, and collect currently the most comprehensive benchmark covering a wide range of scenarios. Extensive experiments on several public datasets and our benchmark demonstrate that our method outputs not only visually appealing fusion but also higher detection mAP than the state-of-the-art approaches.

Most prior work represents the shapes of point clouds by coordinates. However, it is insufficient to describe the local geometry directly. In this paper, we present RepSurf (representative surfaces), a novel representation of point clouds to explicitly depict the very local structure. We explore two variants of RepSurf, Triangular RepSurf and Umbrella RepSurf inspired by triangle meshes and umbrella curvature in computer graphics. We compute the representations of RepSurf by predefined geometric priors after surface reconstruction. RepSurf can be a plug-and-play module for most point cloud models thanks to its free collaboration with irregular points. Based on a simple baseline of PointNet++ (SSG version), Umbrella RepSurf surpasses the previous state-of-the-art by a large margin for classification, segmentation and detection on various benchmarks in terms of performance and efficiency. With an increase of around 0.008M number of parameters, 0.04G FLOPs, and 1.12ms inference time, our method achieves 94.7% (+0.5%) on ModelNet40, and 84.6% (+1.8%) on ScanObjectNN for classification, while 74.3% (+0.8%) mIoU on S3DIS 6-fold, and 70.0% (+1.6%) mIoU on ScanNet for segmentation. For detection, previous state-of-the-art detector with our RepSurf obtains 71.2% (+2.1%) mAP_25, 54.8% (+2.0%) mAP_50 on ScanNetV2, and 64.9% (+1.9%) mAP_25, 47.1% (+2.5%) mAP_50 on SUN RGB-D. Our lightweight Triangular RepSurf performs its excellence on these benchmarks as well. The code is publicly available at https://github.com/hancyran/RepSurf.

Super-Fibonacci spirals are an extension of Fibonacci spirals, enabling fast generation of an arbitrary but fixed number of 3D orientations. The algorithm is simple and fast. A comprehensive evaluation comparing to other methods shows that the generated sets of orientations have low discrepancy, minimal spurious components in the power spectrum, and almost identical Voronoi volumes. This makes them useful for a variety of applications in vision, robotics, machine learning, and in particular Monte Carlo sampling.

Single-step adversarial training (AT) has received wide attention as it proved to be both efficient and robust. However, a serious problem of catastrophic overfitting exists, i.e., the robust accuracy against projected gradient descent (PGD) attack suddenly drops to 0% during the training. In this paper, we approach this problem from a novel perspective of optimization and firstly reveal the close link between the fast-growing gradient of each sample and overfitting, which can also be applied to understand robust overfitting in multi-step AT. To control the growth of the gradient, we propose a new AT method, Subspace Adversarial Training (Sub-AT), which constrains AT in a carefully extracted subspace. It successfully resolves both kinds of overfitting and significantly boosts the robustness. In subspace, we also allow single-step AT with larger steps and larger radius, further improving the robustness performance. As a result, we achieve state-of-the-art single-step AT performance. Without any regularization term, our single-step AT can reach over 51% robust accuracy against strong PGD-50 attack of radius 8/255 on CIFAR-10, reaching a competitive performance against standard multi-step PGD-10 AT with huge computational advantages. The code is released at https://github.com/nblt/Sub-AT.

We introduce a high resolution, 3D-consistent image and shape generation technique which we call StyleSDF. Our method is trained on single view RGB data only, and stands on the shoulders of StyleGAN2 for image generation, while solving two main challenges in 3D-aware GANs: 1) high-resolution, view-consistent generation of the RGB images, and 2) detailed 3D shape. We achieve this by merging an SDF-based 3D representation with a style-based 2D generator. Our 3D implicit network renders low-resolution feature maps, from which the style-based network generates view-consistent, 1024x1024 images. Notably, our SDF-based 3D modeling defines detailed 3D surfaces, leading to consistent volume rendering. Our method shows higher quality results compared to state of the art in terms of visual and geometric quality.

We propose a memory efficient method, named Stochastic Backpropagation (SBP), for training deep neural networks on videos. It is based on the finding that gradients from incomplete execution for backpropagation can still effectively train the models with minimal accuracy loss, which attributes to the high redundancy of video. SBP keeps all forward paths but randomly and independently removes the backward paths for each network layer in each training step. It reduces the GPU memory cost by eliminating the need to cache activation values corresponding to the dropped backward paths, whose amount can be controlled by an adjustable keep-ratio. Experiments show that SBP can be applied to a wide range of models for video tasks, leading to up to 80.0% GPU memory saving and 10% training speedup with less than 1% accuracy drop on action recognition and temporal action detection.

We present a method for semantically transferring the visual appearance of one natural image to another. Specifically, our goal is to generate an image in which objects in a source structure image are "painted" with the visual appearance of their semantically related objects in a target appearance image. Our method works by training a generator given only a single structure/appearance image pair as input. To integrate semantic information into our framework---a pivotal component in tackling this task---our key idea is to leverage a pre-trained and fixed Vision Transformer (ViT) model which serves as an external semantic prior. Specifically, we derive novel representations of structure and appearance extracted from deep ViT features, untwisting them from the learned self-attention modules. We then establish an objective function that splices the desired structure and appearance representations, interweaving them together in the space of ViT features. Our framework, which we term "Splice", does not involve adversarial training, nor does it require any additional input information such as semantic segmentation or correspondences, and can generate high resolution results, e.g., work in HD. We demonstrate high quality results on a variety of in-the-wild image pairs, under significant variations in the number of objects, their pose and appearance.

The 360deg imaging has recently gained much attention; however, its angular resolution is relatively lower than that of a narrow field-of-view (FOV) perspective image as it is captured using a fisheye lens with the same sensor size. Therefore, it is beneficial to super-resolve a 360deg image. Several attempts have been made, but mostly considered equirectangular projection (ERP) as one of the ways for 360deg image representation despite the latitude-dependent distortions. In that case, as the output high-resolution (HR) image is always in the same ERP format as the lowresolution (LR) input, additional information loss may occur when transforming the HR image to other projection types. In this paper, we propose SphereSR, a novel framework to generate a continuous spherical image representation from an LR 360deg image, with the goal of predicting the RGB values at given spherical coordinates for superresolution with an arbitrary 360deg image projection. Specifically, first we propose a feature extraction module that represents the spherical data based on an icosahedron and that efficiently extracts features on the spherical surface. We then propose a spherical local implicit image function (SLIIF) to predict RGB values at the spherical coordinates. As such, SphereSR flexibly reconstructs an HR image given an arbitrary projection type. Experiments on various benchmark datasets show that the proposed method significantly surpasses existing methods in terms of performance.

Current LiDAR-only 3D detection methods inevitably suffer from the sparsity of point clouds. Many multi-modal methods are proposed to alleviate this issue, while different representations of images and point clouds make it difficult to fuse them, resulting in suboptimal performance. In this paper, we present a novel multi-modal framework SFD (Sparse Fuse Dense), which utilizes pseudo point clouds generated from depth completion to tackle the issues mentioned above. Different from prior works, we propose a new RoI fusion strategy 3D-GAF (3D Grid-wise Attentive Fusion) to make fuller use of information from different types of point clouds. Specifically, 3D-GAF fuses 3D RoI features from the pair of point clouds in a grid-wise attentive way, which is more fine-grained and more precise. In addition, we propose a SynAugment (Synchronized Augmentation) to enable our multi-modal framework to utilize all data augmentation approaches tailored to LiDAR-only methods. Lastly, we customize an effective and efficient feature extractor CPConv (Color Point Convolution) for pseudo point clouds. It can explore 2D image features and 3D geometric features of pseudo point clouds simultaneously. Our method holds the highest entry on the KITTI car 3D object detection leaderboard, demonstrating the effectiveness of our SFD. Code will be made publicly available.

Source-free object detection (SFOD) needs to adapt a detector pre-trained on a labeled source domain to a target domain, with only unlabeled training data from the target domain. Existing SFOD methods typically adopt the pseudo labeling paradigm with model adaption alternating between predicting pseudo labels and fine-tuning the model. This approach suffers from both unsatisfactory accuracy of pseudo labels due to the presence of domain shift and limited use of target domain training data. In this work, we present a novel Learning to Overlook Domain Style (LODS) method with such limitations solved in a principled manner. Our idea is to reduce the domain shift effect by enforcing the model to overlook the target domain style, such that model adaptation is simplified and becomes easier to carry on. To that end, we enhance the style of each target domain image and leverage the style degree difference between the original image and the enhanced image as a self-supervised signal for model adaptation. By treating the enhanced image as an auxiliary view, we exploit a student-teacher architecture for learning to overlook the style degree difference against the original image, also characterized with a novel style enhancement algorithm and graph alignment constraint. Extensive experiments demonstrate that our LODS yields new state-of-the-art performance on four benchmarks.

Existing state-of-the-art 3D instance segmentation methods perform semantic segmentation followed by grouping. The hard predictions are made when performing semantic segmentation such that each point is associated with a single class. However, the errors stemming from hard decision propagate into grouping that results in (1) low overlaps between the predicted instance with the ground truth and (2) substantial false positives. To address the aforementioned problems, this paper proposes a 3D instance segmentation method referred to as SoftGroup by performing bottom-up soft grouping followed by top-down refinement. SoftGroup allows each point to be associated with multiple classes to mitigate the problems stemming from semantic prediction errors and suppresses false positive instances by learning to categorize them as background. Experimental results on different datasets and multiple evaluation metrics demonstrate the efficacy of SoftGroup. Its performance surpasses the strongest prior method by a significant margin of +6.2% on the ScanNet v2 hidden test set and +6.8% on S3DIS Area 5 in terms of AP50. SoftGroup is also fast, running at 345ms per scan with a single Titan X on ScanNet v2 dataset. The source code and trained models for both datasets are available at https://github.com/thangvubk/SoftGroup.git.

Face-swapping models have been drawing attention for their compelling generation quality, but their complex architectures and loss functions often require careful tuning for successful training. We propose a new face-swapping model called 'Smooth-Swap', which excludes complex handcrafted designs and allows fast and stable training. The main idea of Smooth-Swap is to build smooth identity embedding that can provide stable gradients for identity change. Unlike the one used in previous models trained for a purely discriminative task, the proposed embedding is trained with a supervised contrastive loss promoting a smoother space. With improved smoothness, Smooth-Swap suffices to be composed of a generic U-Net-based generator and three basic loss functions, a far simpler design compared with the previous models. Extensive experiments on face-swapping benchmarks (FFHQ, FaceForensics++) and face images in the wild show that our model is also quantitatively and qualitatively comparable or even superior to the existing methods.

Recent Vision Transformer (ViT) models have demonstrated encouraging results across various computer vision tasks, thanks to its competence in modeling long-range dependencies of image patches or tokens via self-attention. These models, however, usually designate the similar receptive fields of each token feature within each layer. Such a constraint inevitably limits the ability of each self-attention layer in capturing multi-scale features, thereby leading to performance degradation in handling images with multiple objects of different scales. To address this issue, we propose a novel and generic strategy, termed shunted self-attention (SSA), that allows ViTs to model the attentions at hybrid scales per attention layer. The key idea of SSA is to inject heterogeneous receptive field sizes into tokens: before computing the self-attention matrix, it selectively merges tokens to represent larger object features while keeping certain tokens to preserve fine-grained features. This novel merging scheme enables the self-attention to learn relationships between objects with different sizes and simultaneously reduces the token numbers and the computational cost. Extensive experiments across various tasks demonstrate the superiority of SSA. Specifically, the SSA-based transformer achieves 84.0% Top-1 accuracy and outperforms the state-of-the-art Focal Transformer on ImageNet with only half of the model size and computation cost, and surpasses Focal Transformer by 1.3 mAP on COCO and 2.9 mIOU on ADE20K under similar parameter and computation cost. Code has been released at \href https://github.com/OliverRensu/Shunted-Transformer https://github.com/OliverRensu/Shunted-Transformer .

Deep models trained on source domain lack generalization when evaluated on unseen target domains with different data distributions. The problem becomes even more pronounced when we have no access to target domain samples for adaptation. In this paper, we address domain generalized semantic segmentation, where a segmentation model is trained to be domain-invariant without using any target domain data. Existing approaches to tackle this problem standardize data into a unified distribution. We argue that while such a standardization promotes global normalization, the resulting features are not discriminative enough to get clear segmentation boundaries. To enhance separation between categories while simultaneously promoting domain invariance, we propose a framework including two novel modules: Semantic-Aware Normalization (SAN) and Semantic-Aware Whitening (SAW). Specifically, SAN focuses on category-level center alignment between features from different image styles, while SAW enforces distributed alignment for the already center-aligned features. With the help of SAN and SAW, we encourage both intraclass compactness and inter-class separability. We validate our approach through extensive experiments on widely-used datasets (i.e. GTAV, SYNTHIA, Cityscapes, Mapillary and BDDS). Our approach shows significant improvements over existing state-of-the-art on various backbone networks. Code is available at https://github.com/leolyj/SAN-SAW

We propose SelfRecon, a clothed human body reconstruction method that combines implicit and explicit representations to recover space-time coherent geometries from a monocular self-rotating human video. Explicit methods require a predefined template mesh for a given sequence, while the template is hard to acquire for a specific subject. Meanwhile, the fixed topology limits the reconstruction accuracy and clothing types. Implicit representation supports arbitrary topology and can represent high-fidelity geometry shapes due to its continuous nature. However, it is difficult to integrate multi-frame information to produce a consistent registration sequence for downstream applications. We propose to combine the advantages of both representations. We utilize differential mask loss of the explicit mesh to obtain the coherent overall shape, while the details on the implicit surface are refined with the differentiable neural rendering. Meanwhile, the explicit mesh is updated periodically to adjust its topology changes, and a consistency loss is designed to match both representations. Compared with existing methods, SelfRecon can produce high-fidelity surfaces for arbitrary clothed humans with self-supervised optimization. Extensive experimental results demonstrate its effectiveness on real captured monocular videos. The source code is available at https://github.com/jby1993/SelfReconCode.

In this paper, we propose self-supervised training for video transformers using unlabeled video data. From a given video, we create local and global spatiotemporal views with varying spatial sizes and frame rates. Our self-supervised objective seeks to match the features of these different views representing the same video, to be invariant to spatiotemporal variations in actions. To the best of our knowledge, the proposed approach is the first to alleviate the dependency on negative samples or dedicated memory banks in Self-supervised Video Transformer (SVT). Further, owing to the flexibility of Transformer models, SVT supports slow-fast video processing within a single architecture using dynamically adjusted positional encoding and supports long-term relationship modeling along spatiotemporal dimensions. Our approach performs well on four action recognition benchmarks (Kinetics-400, UCF-101, HMDB-51, and SSv2) and converges faster with small batch sizes. Code is available at: https://git.io/J1juJ