CVPR 2021 论文列表

Scene flow depicts the dynamics of a 3D scene, which is critical for various applications such as autonomous driving, robot navigation, AR/VR, etc. Conventionally, scene flow is estimated from dense/regular RGB video frames. With the development of depth-sensing technologies, precise 3D measurements are available via point clouds which have sparked new research in 3D scene flow. Nevertheless, it remains challenging to extract scene flow from point clouds due to the sparsity and irregularity in typical point cloud sampling patterns. One major issue related to irregular sampling is identified as the randomness during point set abstraction/feature extraction---an elementary process in many flow estimation scenarios. A novel Spatial Abstraction with Attention (SA^2) layer is accordingly proposed to alleviate the unstable abstraction problem. Moreover, a Temporal Abstraction with Attention (TA^2) layer is proposed to rectify attention in temporal domain, leading to benefits with motions scaled in a larger range. Extensive analysis and experiments verified the motivation and significant performance gains of our method, dubbed as Flow Estimation via Spatial-Temporal Attention (FESTA), when compared to several state-of-the-art benchmarks of scene flow estimation.

We propose a fully convolutional multi-person pose estimation framework using dynamic instance-aware convolutions, termed FCPose. Different from existing methods, which often require ROI (Region of Interest) operations and/or grouping post-processing, FCPose eliminates the ROIs and grouping post-processing with dynamic instance-aware keypoint estimation heads. The dynamic keypoint heads are conditioned on each instance (person), and can encode the instance concept in the dynamically-generated weights of their filters. Moreover, with the strong representation capacity of dynamic convolutions, the keypoint heads in FCPose are designed to be very compact, resulting in fast inference and makes FCPose have almost constant inference time regardless of the number of persons in the image. For example, on the COCO dataset, a real-time version of FCPose using the DLA-34 backbone infers about 4.5 times faster than Mask R-CNN (ResNet-101) (41.67 FPS vs. 9.26 FPS) while achieving improved performance (64.8% AP vs. 64.3% AP). FCPose also offers better speed/accuracy trade-off than other state-of-the-art methods. Our experiment results show that FCPose is a simple yet effective multi-person pose estimation framework. Code is available at: https://git.io/AdelaiDet

Neural Architecture Search (NAS) yields state-of-the-art neural networks that outperform their best manually-designed counterparts. However, previous NAS methods search for architectures under one set of training hyper-parameters (i.e., a training recipe), overlooking superior architecture-recipe combinations. To address this, we present Neural Architecture-Recipe Search (NARS) to search both (a) architectures and (b) their corresponding training recipes, simultaneously. NARS utilizes an accuracy predictor that scores architecture and training recipes jointly, guiding both sample selection and ranking. Furthermore, to compensate for the enlarged search space, we leverage "free" architecture statistics (e.g., FLOP count) to pretrain the predictor, significantly improving its sample efficiency and prediction reliability. After training the predictor via constrained iterative optimization, we run fast evolutionary searches in just CPU minutes to generate architecture-recipe pairs for a variety of resource constraints, called FBNetV3. FBNetV3 makes up a family of state-of-the-art compact neural networks that outperform both automatically and manually-designed competitors. For example, FBNetV3 matches both EfficientNet and ResNeSt accuracy on ImageNet with up to 2.0x and 7.1x fewer FLOPs, respectively. Furthermore, FBNetV3 yields significant performance gains for downstream object detection tasks, improving mAP despite 18% fewer FLOPs and 34% fewer parameters than EfficientNet-based equivalents.

We consider the challenging blind denoising problem for Poisson-Gaussian noise, in which no additional information about clean images or noise level parameters is available. Particularly, when only "single" noisy images are available for training a denoiser, the denoising performance of existing methods was not satisfactory. Recently, the blind pixelwise affine image denoiser (BP-AIDE) was proposed and significantly improved the performance in the above setting, to the extent that it is competitive with denoisers which utilized additional information. However, BP-AIDE seriously suffered from slow inference time due to the inefficiency of noise level estimation procedure and that of the blind-spot network (BSN) architecture it used. To that end, we propose Fast Blind Image Denoiser (FBI-Denoiser) for Poisson-Gaussian noise, which consists of two neural network models; 1) PGE-Net that estimates Poisson-Gaussian noise parameters 2000 times faster than the conventional methods and 2) FBI-Net that realizes a much more efficient BSN for pixelwise affine denoiser in terms of the number of parameters and inference speed. Consequently, we show that our FBI-Denoiser blindly trained solely based on single noisy images can achieve the state-of-the-art performance on several real-world noisy image benchmark datasets with much faster inference time (X 10), compared to BP-AIDE.

This paper is about few-shot instance segmentation, where training and test image sets do not share the same object classes. We specify and evaluate a new few-shot anchor-free part-based instance segmenter (FAPIS). Our key novelty is in explicit modeling of latent object parts shared across training object classes, which is expected to facilitate our few-shot learning on new classes in testing. We specify a new anchor-free object detector aimed at scoring and regressing locations of foreground bounding boxes, as well as estimating relative importance of latent parts within each box. Also, we specify a new network for delineating and weighting latent parts for the final instance segmentation within every detected bounding box. Our evaluation on the benchmark COCO-20i dataset demonstrates that we significantly outperform the state of the art.

Image caption evaluation is a crucial task, which involves the semantic perception and matching of image and text. Good evaluation metrics aim to be fair, comprehensive, and consistent with human judge intentions. When humans evaluate a caption, they usually consider multiple aspects, such as whether it is related to the target image without distortion, how much image gist it conveys, as well as how fluent and beautiful the language and wording is. The above three different evaluation orientations can be summarized as fidelity, adequacy, and fluency. The former two rely on the image content, while fluency is purely related to linguistics and more subjective. Inspired by human judges, we propose a learning-based metric named FAIEr to ensure evaluating the fidelity and adequacy of the captions. Since image captioning involves two different modalities, we employ the scene graph as a bridge between them to represent both images and captions. FAIEr mainly regards the visual scene graph as the criterion to measure the fidelity. Then for evaluating the adequacy of the candidate caption, it highlights the image gist on the visual scene graph under the guidance of the reference captions. Comprehensive experimental results show that FAIEr has high consistency with human judgment as well as high stability, low reference dependency, and the capability of reference-free evaluation.

We present a technique for estimating the relative 3D rotation of an RGB image pair in an extreme setting, where the images have little or no overlap. We observe that, even when images do not overlap, there may be rich hidden cues as to their geometric relationship, such as light source directions, vanishing points, and symmetries present in the scene. We propose a network design that can automatically learn such implicit cues by comparing all pairs of points between the two input images. Our method therefore constructs dense feature correlation volumes and processes these to predict relative 3D rotations. Our predictions are formed over a fine-grained discretization of rotations, bypassing difficulties associated with regressing 3D rotations. We demonstrate our approach on a large variety of extreme RGB image pairs, including indoor and outdoor images captured under different lighting conditions and geographic locations. Our evaluation shows that our model can successfully estimate relative rotations among non-overlapping images without compromising performance over overlapping image pairs.

Recently, learning-based approaches have achieved impressive results in the field of low-light image denoising. Some state of the art approaches employ a rich physical model to generate realistic training data. However, the performance of these approaches ultimately depends on the realism of the physical model, and many works only concentrate on everyday photography. In this work we present a denoising approach for extremely low-light images of permanently shadowed regions (PSRs) on the lunar surface, taken by the Narrow Angle Camera on board the Lunar Reconnaissance Orbiter satellite. Our approach extends existing learning-based approaches by combining a physical noise model of the camera with real noise samples and training image scene selection based on 3D ray tracing to generate realistic training data. We also condition our denoising model on the camera's environmental metadata at the time of image capture (such as the camera's temperature and age), showing that this improves performance. Our quantitative and qualitative results show that our method strongly outperforms the existing calibration routine for the camera and other baselines. Our results could significantly impact lunar science and exploration, for example by aiding the identification of surface water-ice and reducing uncertainty in rover and human traverse planning into PSRs.

We present a plug-in replacement for batch normalization (BN) called exponential moving average normalization (EMAN), which improves the performance of existing student-teacher based self- and semi-supervised learning techniques. Unlike the standard BN, where the statistics are computed within each batch, EMAN, used in the teacher, updates its statistics by exponential moving average from the BN statistics of the student. This design reduces the intrinsic cross-sample dependency of BN and enhances the generalization of the teacher. EMAN improves strong baselines for self-supervised learning by 4-6/1-2 points and semi-supervised learning by about 7/2 points, when 1%/10% supervised labels are available on ImageNet. These improvements are consistent across methods, network architectures, training duration, and datasets, demonstrating the general effectiveness of this technique. The code will be made available online.

We present a new learning-based framework to recover vehicle pose in SO(3) from a single RGB image. In contrast to previous works that map local appearance to observation angles, we explore a progressive approach by extracting meaningful Intermediate Geometrical Representations (IGRs) to estimate egocentric vehicle orientation. This approach features a deep model that transforms perceived intensities to IGRs, which are mapped to a 3D representation encoding object orientation in the camera coordinate system. Core problems are what IGRs to use and how to learn them more effectively. We answer the former question by designing IGRs based on an interpolated cuboid that derives from primitive 3D annotation readily. The latter question motivates us to incorporate geometry knowledge with a new loss function based on a projective invariant. This loss function allows unlabeled data to be used in the training stage to improve representation learning. Without additional labels, our system outperforms previous monocular RGB-based methods for joint vehicle detection and pose estimation on the KITTI benchmark, achieving performance even comparable to stereo methods. Code and pre-trained models are available at this HTTPS URL.

Automatically generating radiology reports can improve current clinical practice in diagnostic radiology. On one hand, it can relieve radiologists from the heavy burden of report writing; On the other hand, it can remind radiologists of abnormalities and avoid the misdiagnosis and missed diagnosis. Yet, this task remains a challenging job for data-driven neural networks, due to the serious visual and textual data biases. To this end, we propose a Posterior-and-Prior Knowledge Exploring-and-Distilling approach (PPKED) to imitate the working patterns of radiologists, who will first examine the abnormal regions and assign the disease topic tags to the abnormal regions, and then rely on the years of prior medical knowledge and prior working experience accumulations to write reports. Thus, the PPKED includes three modules: Posterior Knowledge Explorer (PoKE), Prior Knowledge Explorer (PrKE) and Multi-domain Knowledge Distiller (MKD). In detail, PoKE explores the posterior knowledge, which provides explicit abnormal visual regions to alleviate visual data bias; PrKE explores the prior knowledge from the prior medical knowledge graph (medical knowledge) and prior radiology reports (working experience) to alleviate textual data bias. The explored knowledge is distilled by the MKD to generate the final reports. Evaluated on MIMIC-CXR and IU-Xray datasets, our method is able to outperform previous state-of-the-art models on these two datasets.

Current CNN-based super-resolution (SR) methods process all locations equally with computational resources being uniformly assigned in space. However, since missing details in low-resolution (LR) images mainly exist in regions of edges and textures, less computational resources are required for those flat regions. Therefore, existing CNN-based methods involve redundant computation in flat regions, which increases their computational cost and limits their applications on mobile devices. In this paper, we explore the sparsity in image SR to improve inference efficiency of SR networks. Specifically, we develop a Sparse Mask SR (SMSR) network to learn sparse masks to prune redundant computation. Within our SMSR, spatial masks learn to identify "important" regions while channel masks learn to mark redundant channels in those "unimportant" regions. Consequently, redundant computation can be accurately localized and skipped while maintaining comparable performance. It is demonstrated that our SMSR achieves state-of-the-art performance with 41%/33%/27% FLOPs being reduced for x2/3/4 SR. Code is available at: https://github.com/LongguangWang/SMSR.

Siamese networks have become a common structure in various recent models for unsupervised visual representation learning. These models maximize the similarity between two augmentations of one image, subject to certain conditions for avoiding collapsing solutions. In this paper, we report surprising empirical results that simple Siamese networks can learn meaningful representations even using none of the following: (i) negative sample pairs, (ii) large batches, (iii) momentum encoders. Our experiments show that collapsing solutions do exist for the loss and structure, but a stop-gradient operation plays an essential role in preventing collapsing. We provide a hypothesis on the implication of stop-gradient, and further show proof-of-concept experiments verifying it. Our "SimSiam" method achieves competitive results on ImageNet and downstream tasks. We hope this simple baseline will motivate people to rethink the roles of Siamese architectures for unsupervised representation learning. Code is made available. (https://github.com/facebookresearch/simsiam)

We investigate the weakly-supervised audio-visual video parsing task, which aims to parse a video into temporal event segments and predict the audible or visible event categories. The task is challenging since there only exist video-level event labels for training, without indicating the temporal boundaries and modalities. Previous works take the overall event labels to supervise both audio and visual model predictions. However, we argue that such overall labels harm the model training due to the audio-visual asynchrony. For example, commentators speak in a basketball video, but we cannot visually find the speakers. In this paper, we tackle this issue by leveraging the cross-modal correspondence of audio and visual signals. We generate reliable event labels individually for each modality by swapping audio and visual tracks with other unrelated videos. If the original visual/audio data contain event clues, the event prediction from the newly assembled data would still be highly confident. In this way, we could protect our models from being misled by ambiguous event labels. In addition, we propose the cross-modal audio-visual contrastive learning to induce temporal difference on attention models within videos, i.e., urging the model to pick the current temporal segment from all context candidates. Experiments show we outperform state-of-the-art methods by a large margin.

The rapid progress in 3D scene understanding has come with growing demand for data; however, collecting and annotating 3D scenes (e.g. point clouds) are notoriously hard. For example, the number of scenes (e.g. indoor rooms) that can be accessed and scanned might be limited; even given sufficient data, acquiring 3D labels (e.g. instance masks) requires intensive human labor. In this paper, we explore data-efficient learning for 3D point cloud. As a first step towards this direction, we propose Contrastive Scene Contexts, a 3D pre-training method that makes use of both point-level correspondences and spatial contexts in a scene. Our method achieves state-of-the-art results on a suite of benchmarks where training data or labels are scarce. Our study reveals that exhaustive labelling of 3D point clouds might be unnecessary; and remarkably, on ScanNet, even using 0.1% of point labels, we still achieve 89% (instance segmentation) and 96% (semantic segmentation) of the baseline performance that uses full annotations.

In many real-world problems, collecting a large number of labeled samples is infeasible. Few-shot learning (FSL) is the dominant approach to address this issue, where the objective is to quickly adapt to novel categories in presence of a limited number of samples. FSL tasks have been predominantly solved by leveraging the ideas from gradient-based meta-learning and metric learning approaches. However, recent works have demonstrated the significance of powerful feature representations with a simple embedding network that can outperform existing sophisticated FSL algorithms. In this work, we build on this insight and propose a novel training mechanism that simultaneously enforces equivariance and invariance to a general set of geometric transformations. Equivariance or invariance has been employed standalone in the previous works; however, to the best of our knowledge, they have not been used jointly. Simultaneous optimization for both of these contrasting objectives allows the model to jointly learn features that are not only independent of the input transformation but also the features that encode the structure of geometric transformations. These complementary sets of features help generalize well to novel classes with only a few data samples. We achieve additional improvements by incorporating a novel self-supervised distillation objective. Our extensive experimentation shows that even without knowledge distillation our proposed method can outperform current state-of-the-art FSL methods on five popular benchmark datasets.

Images synthesized by powerful generative adversarial network (GAN) based methods have drawn moral and privacy concerns. Although image forensic models have reached great performance in detecting fake images from real ones, these models can be easily fooled with a simple adversarial attack. But, the noise adding adversarial samples are also arousing suspicion. In this paper, instead of adding adversarial noise, we optimally search adversarial points on face manifold to generate anti-forensic fake face images. We iteratively do a gradient-descent with each small step in the latent space of a generative model, e.g. Style-GAN, to find an adversarial latent vector, which is similar to norm-based adversarial attack but in latent space. Then, the generated fake images driven by the adversarial latent vectors with the help of GANs can defeat main-stream forensic models. For examples, they make the accuracy of deepfake detection models based on Xception or EfficientNet drop from over 90% to nearly 0%, meanwhile maintaining high visual quality. In addition, we find manipulating noise vectors n at different levels have different impacts on attack success rate, and the generated adversarial images mainly have changes on facial texture or face attributes.

This paper introduces a method to encode the blur operators of an arbitrary dataset of sharp-blur image pairs into a blur kernel space. Assuming the encoded kernel space is close enough to in-the-wild blur operators, we propose an alternating optimization algorithm for blind image deblurring. It approximates an unseen blur operator by a kernel in the encoded space and searches for the corresponding sharp image. Unlike recent deep-learning-based methods, our system can handle unseen blur kernel, while avoiding using complicated handcrafted priors on the blur operator often found in classical methods. Due to the method's design, the encoded kernel space is fully differentiable, thus can be easily adopted in deep neural network models. Moreover, our method can be used for blur synthesis by transferring existing blur operators from a given dataset into a new domain. Finally, we provide experimental results to confirm the effectiveness of the proposed method.

Generative adversarial networks (GANs) synthesize realistic images from random latent vectors. Although manipulating the latent vectors controls the synthesized outputs, editing real images with GANs suffers from i) time-consuming optimization for projecting real images to the latent vectors, ii) or inaccurate embedding through an encoder. We propose StyleMapGAN: the intermediate latent space has spatial dimensions, and a spatially variant modulation replaces AdaIN. It makes the embedding through an encoder more accurate than existing optimization-based methods while maintaining the properties of GANs. Experimental results demonstrate that our method significantly outperforms state-of-the-art models in various image manipulation tasks such as local editing and image interpolation. Last but not least, conventional editing methods on GANs are still valid on our StyleMapGAN. Source code is available at https://github.com/naver-ai/StyleMapGAN.