CVPR 2021 论文列表

Face clustering is a promising method for annotating unlabeled face images. Recent supervised approaches have boosted the face clustering accuracy greatly, however their performance is still far from satisfactory. These methods can be roughly divided into global-based and local-based ones. Global-based methods suffer from the limitation of training data scale, while local-based ones are difficult to grasp the whole graph structure information and usually take a long time for inference. Previous approaches fail to tackle these two challenges simultaneously. To address the dilemma of large-scale training and efficient inference, we propose the STructure-AwaRe Face Clustering (STAR-FC) method. Specifically, we design a structure-preserved subgraph sampling strategy to explore the power of large-scale training data, which can increase the training data scale from 10^5 to 10^7. During inference, the STAR-FC performs efficient full-graph clustering with two steps: graph parsing and graph refinement. And the concept of node intimacy is introduced in the second step to mine the local structural information. The STAR-FC gets 91.97 pairwise F-score on partial MS1M within 310s which surpasses the state-of-the-arts. Furthermore, we are the first to train on very large-scale graph with 20M nodes, and achieve superior inference results on 12M testing data. Overall, as a simple and effective method, the proposed STAR-FC provides a strong baseline for large-scale face clustering. Code is available at https://sstzal.github.io/STAR-FC/.

Monocular 3D prediction is one of the fundamental problems in 3D vision. Recent deep learning-based approaches have brought us exciting progress on this problem. However, existing approaches have predominantly focused on end-to-end depth and normal predictions, which do not fully utilize the underlying 3D environment's geometric structures. This paper introduces StruMonoNet, which detects and enforces a planar structure to enhance pixel-wise predictions. StruMonoNet innovates in leveraging a hybrid representation that combines visual feature and a surfel representation for plane prediction. This formulation allows us to combine the power of visual feature learning and the flexibility of geometric representations in incorporating geometric relations. As a result, StruMonoNet can detect relations between planes such as adjacent planes, perpendicular planes, and parallel planes, all of which are beneficial for dense 3D prediction. Experimental results show that StruMonoNet considerably outperforms state-of-the-art approaches on NYUv2 and ScanNet.

Weakly supervised object localization (WSOL) aims at learning to localize objects of interest by only using the image-level labels as the supervision. While numerous efforts have been made in this field, recent approaches still suffer from two challenges: one is the part domination issue while the other is the learning robustness issue. Specifically, the former makes the localizer prone to the local discriminative object regions rather than the desired whole object, and the latter makes the localizer over-sensitive to the variations of the input images so that one can hardly obtain localization results robust to the arbitrary visual stimulus. To solve these issues, we propose a novel framework to strengthen the learning tolerance, referred to as SLT-Net, for WSOL. Specifically, we consider two-fold learning tolerance strengthening mechanisms. One is the semantic tolerance strengthening mechanism, which allows the localizer to make mistakes for classifying similar semantics so that it will not concentrate too much on the discriminative local regions. The other is the visual stimuli tolerance strengthening mechanism, which enforces the localizer to be robust to different image transformations so that the prediction quality will not be sensitive to each specific input image. Finally, we implement comprehensive experimental comparisons on two widely-used datasets CUB and ILSVRC2012, which demonstrate the effectiveness of our proposed approach.

Batch Normalization (BN) is a popular technique for training Deep Neural Networks (DNNs). BN uses scaling and shifting to normalize activations of mini-batches to accelerate convergence and improve generalization. The recently proposed Iterative Normalization (IterNorm) method improves these properties by whitening the activations iteratively using Newton's method. However, since Newton's method initializes the whitening matrix independently at each training step, no information is shared between consecutive steps. In this work, instead of exact computation of whitening matrix at each time step, we estimate it gradually during training in an online fashion, using our proposed Stochastic Whitening Batch Normalization (SWBN) algorithm. We show that while SWBN improves the convergence rate and generalization of DNNs, its computational overhead is less than that of IterNorm. Due to the high efficiency of the proposed method, it can be easily employed in most DNN architectures with a large number of layers. We provide comprehensive experiments and comparisons between BN, IterNorm, and SWBN layers to demonstrate the effectiveness of the proposed technique in conventional (many-shot) image classification and few-shot classification tasks.

Video understanding calls for a model to learn the characteristic interplay between static scene content and its dynamics: Given an image, the model must be able to predict a future progression of the portrayed scene and, conversely, a video should be explained in terms of its static image content and all the remaining characteristics not present in the initial frame. This naturally suggests a bijective mapping between the video domain and the static content as well as residual information. In contrast to common stochastic image-to-video synthesis, such a model does not merely generate arbitrary videos progressing the initial image. Given this image, it rather provides a one-to-one mapping between the residual vectors and the video with stochastic outcomes when sampling. The approach is naturally implemented using a conditional invertible neural network (cINN) that can explain videos by independently modelling static and other video characteristics, thus laying the basis for controlled video synthesis. Experiments on diverse video datasets demonstrate the effectiveness of our approach in terms of both the quality and diversity of the synthesized results. Our project page is available at https://bit.ly/3dg90fV.

Robust point cloud registration in real-time is an important prerequisite for many mapping and localization algorithms. Traditional methods like ICP tend to fail without good initialization, insufficient overlap or in the presence of dynamic objects. Modern deep learning based registration approaches present much better results, but suffer from a heavy runtime. We overcome these drawbacks by introducing StickyPillars, a fast, accurate and extremely robust deep middle-end 3D feature matching method on point clouds. It uses graph neural networks and performs context aggregation on sparse 3D key-points with the aid of transformer based multi-head self and cross-attention. The network output is used as the cost for an optimal transport problem whose solution yields the final matching probabilities. The system does not rely on hand crafted feature descriptors or heuristic matching strategies. We present state-of-art art accuracy results on the registration problem demonstrated on the KITTI dataset while being four times faster then leading deep methods. Furthermore, we integrate our matching system into a LiDAR odometry pipeline yielding most accurate results on the KITTI odometry dataset. Finally, we demonstrate robustness on KITTI odometry. Our method remains stable in accuracy where state-of-the-art procedures fail on frame drops and higher speeds.

In this paper, we propose StereoPIFu, which integrates the geometric constraints of stereo vision with implicit function representation of PIFu, to recover the 3D shape of the clothed human from a pair of low-cost rectified images. First, we introduce the effective voxel-aligned features from a stereo vision-based network to enable depth-aware reconstruction. Moreover, the novel relative z-offset is employed to associate predicted high-fidelity human depth and occupancy inference, which helps restore fine-level surface details. Second, a network structure that fully utilizes the geometry information from the stereo images is designed to improve the human body reconstruction quality. Consequently, our StereoPIFu can naturally infer the human body's spatial location in camera space and maintain the correct relative position of different parts of the human body, which enables our method to capture human performance. Compared with previous works, our StereoPIFu significantly improves the robustness, completeness, and accuracy of the clothed human reconstruction, which is demonstrated by extensive experimental results.

Recent neural view synthesis methods have achieved impressive quality and realism, surpassing classical pipelines which rely on multi-view reconstruction. State-of-the-Art methods, such as NeRF, are designed to learn a single scene with a neural network and require dense multi-view inputs. Testing on a new scene requires re-training from scratch, which takes 2-3 days. In this work, we introduce Stereo Radiance Fields (SRF), a neural view synthesis approach that is trained end-to-end, generalizes to new scenes, and requires only sparse views at test time. The core idea is a neural architecture inspired by classical multi-view stereo methods, which estimates surface points by finding similar image regions in stereo images. In SRF, we predict color and density for each 3D point given an encoding of its stereo correspondence in the input images. The encoding is implicitly learned by an ensemble of pair-wise similarities -- emulating classical stereo. Experiments show that SRF learns structure instead of overfitting on a scene. We train on multiple scenes of the DTU dataset and generalize to new ones without re-training, requiring only 10 sparse and spread-out views as input. We show that 10-15 minutes of fine-tuning further improve the results, achieving significantly sharper, more detailed results than scene-specific models. The code, model, and videos are available at https://virtualhumans.mpi-inf.mpg.de/srf/.

In applications such as optical see-through and projector augmented reality, producing images amounts to solving non-negative image generation, where one can only add light to an existing image. Most image generation methods, however, are ill-suited to this problem setting, as they make the assumption that one can assign arbitrary color to each pixel. In fact, naive application of existing methods fails even in simple domains such as MNIST digits, since one cannot create darker pixels by adding light. We know, however, that the human visual system can be fooled by optical illusions involving certain spatial configurations of brightness and contrast. Our key insight is that one can leverage this behavior to produce high quality images with negligible artifacts. For example, we can create the illusion of darker patches by brightening surrounding pixels. We propose a novel optimization procedure to produce images that satisfy both semantic and non-negativity constraints. Our approach can incorporate existing state-of-the-art methods, and exhibits strong performance in a variety of tasks including image-to-image translation and style transfer.

We introduce the concept of geometric stability to the problem of 6D object pose estimation and propose to learn pose inference based on geometrically stable patches extracted from observed 3D point clouds. According to the theory of geometric stability analysis, a minimal set of three planar/cylindrical patches are geometrically stable and determine the full 6DoFs of the object pose. We train a deep neural network to regress 6D object pose based on geometrically stable patch groups via learning both intra-patch geometric features and inter-patch contextual features. A subnetwork is jointly trained to predict per-patch poses. This auxiliary task is a relaxation of the group pose prediction: A single patch cannot determine the full 6DoFs but is able to improve pose accuracy in its corresponding DoFs. Working with patch groups makes our method generalize well for random occlusion and unseen instances. The method is easily amenable to resolve symmetry ambiguities. Our method achieves the state-of-the-art results on public benchmarks compared not only to depth-only but also to RGBD methods. It also performs well in category-level pose estimation.

We present Stable View Synthesis (SVS). Given a set of source images depicting a scene from freely distributed viewpoints, SVS synthesizes new views of the scene. The method operates on a geometric scaffold computed via structure-from-motion and multi-view stereo. Each point on this 3D scaffold is associated with view rays and corresponding feature vectors that encode the appearance of this point in the input images. The core of SVS is view-dependent on-surface feature aggregation, in which directional feature vectors at each 3D point are processed to produce a new feature vector for a ray that maps this point into the new target view. The target view is then rendered by a convolutional network from a tensor of features synthesized in this way for all pixels. The method is composed of differentiable modules and is trained end-to-end. It supports spatially-varying view-dependent importance weighting and feature transformation of source images at each point; spatial and temporal stability due to the smooth dependence of on-surface feature aggregation on the target view; and synthesis of view-dependent effects such as specular reflection. Experimental results demonstrate that SVS outperforms state-of-the-art view synthesis methods both quantitatively and qualitatively on three diverse real-world datasets, achieving unprecedented levels of realism in free-viewpoint video of challenging large-scale scenes. Code is available at https://github.com/intel-isl/StableViewSynthesis

We propose a novel approach for multi-modal Image-to-image (I2I) translation. To tackle the one-to-many relationship between input and output domains, previous works use complex training objectives to learn a latent embedding, jointly with the generator, that models the variability of the output domain. In contrast, we directly model the style variability of images, independent of the image synthesis task. Specifically, we pre-train a generic style encoder using a novel proxy task to learn an embedding of images, from arbitrary domains, into a low-dimensional style latent space. The learned latent space introduces several advantages over previous traditional approaches to multi-modal I2I translation. First, it is not dependent on the target dataset, and generalizes well across multiple domains. Second, it learns a more powerful and expressive latent space, which improves the fidelity of style capture and transfer. The proposed style pre-training also simplifies the training objective and speeds up the training significantly. Furthermore, we provide a detailed study of the contribution of different loss terms to the task of multi-modal I2I translation, and propose a simple alternative to VAEs to enable sampling from unconstrained latent spaces. Finally, we achieve state-of-the-art results on six challenging benchmarks with a simple training objective that includes only a GAN loss and a reconstruction loss.

We propose a new formulation for the bundle adjustment problem which relies on nullspace marginalization of landmark variables by QR decomposition. Our approach, which we call square root bundle adjustment, is algebraically equivalent to the commonly used Schur complement trick, improves the numeric stability of computations, and allows for solving large-scale bundle adjustment problems with single-precision floating-point numbers. We show in real-world experiments with the BAL datasets that even in single precision the proposed solver achieves on average equally accurate solutions compared to Schur complement solvers using double precision. It runs significantly faster, but can require larger amounts of memory on dense problems. The proposed formulation relies on simple linear algebra operations and opens the way for efficient implementations of bundle adjustment on hardware platforms optimized for single-precision linear algebra processing.

When people observe events, they are able to abstract key information and build concise summaries of what is happening. These summaries include contextual and semantic information describing the important high-level details (what, where, who and how) of the observed event and exclude background information that is deemed unimportant to the observer. With this in mind, the descriptions people generate for videos of different dynamic events can greatly improve our understanding of the key information of interest in each video. These descriptions can be captured in captions that provide expanded attributes for video labeling (e.g. actions/objects/scenes/sentiment/etc.) while allowing us to gain new insight into what people find important or necessary to summarize specific events. Existing caption datasets for video understanding are either small in scale or restricted to a specific domain. To address this, we present the Spoken Moments (S-MiT) dataset of 500k spoken captions each attributed to a unique short video depicting a broad range of different events. We collect our descriptions using audio recordings to ensure that they remain as natural and concise as possible while allowing us to scale the size of a large classification dataset. In order to utilize our proposed dataset, we present a novel Adaptive Mean Margin (AMM) approach to contrastive learning and evaluate our models on video/caption retrieval on multiple datasets. We show that our AMM approach consistently improves our results and that models trained on our Spoken Moments dataset generalize better than those trained on other video-caption datasets.

The recently invented retina-inspired spike camera has shown great potential for capturing dynamic scenes. Different from the conventional digital cameras that compact the photoelectric information within the exposure interval into a single snapshot, the spike camera produces a continuous spike stream to record the dynamic light intensity variation process. For spike cameras, image reconstruction remains an important and challenging issue. To this end, this paper develops a spike-to-image neural network (Spk2ImgNet) to reconstruct the dynamic scene from the continuous spike stream. In particular, to handle the challenges brought by both noise and high-speed motion, we propose a hierarchical architecture to exploit the temporal correlation of the spike stream progressively. Firstly, a spatially adaptive light inference subnet is proposed to exploit the local temporal correlation, producing basic light intensity estimates of different moments. Then, a pyramid deformable alignment is utilized to align the intermediate features such that the feature fusion module can exploit the long-term temporal correlation, while avoiding undesired motion blur. In addition, to train the network, we simulate the working mechanism of spike camera to generate a large-scale spike dataset composed of spike streams and corresponding ground truth images. Experimental results demonstrate that the proposed network evidently outperforms the state-of-the-art spike camera reconstruction methods.

Extracting robust and general 3D local features is key to downstream tasks such as point cloud registration and reconstruction. Existing learning-based local descriptors are either sensitive to rotation transformations, or rely on classical handcrafted features which are neither general nor representative. In this paper, we introduce a new, yet conceptually simple, neural architecture, termed SpinNet, to extract local features which are rotationally invariant whilst sufficiently informative to enable accurate registration. A Spatial Point Transformer is first introduced to map the input local surface into a carefully designed cylindrical space, enabling end-to-end optimization with SO(2) equivariant representation. A Neural Feature Extractor which leverages the powerful point-based and 3D cylindrical convolutional neural layers is then utilized to derive a compact and representative descriptor for matching. Extensive experiments on both indoor and outdoor datasets demonstrate that SpinNet outperforms existing state-of-the-art techniques by a large margin. More critically, it has the best generalization ability across unseen scenarios with different sensor modalities.

An emerging line of research has found that spherical spaces better match the underlying geometry of facial images, as evidenced by the state-of-the-art facial recognition methods which benefit empirically from spherical representations. Yet, these approaches rely on deterministic embeddings and hence suffer from the feature ambiguity dilemma, whereby ambiguous or noisy images are mapped into poorly learned regions of representation space, leading to inaccuracies. Probabilistic Face Embeddings (PFE) is the first attempt to address this dilemma. However, we theoretically and empirically identify two main failures of PFE when it is applied to spherical deterministic embeddings aforementioned. To address these issues, in this paper, we propose a novel framework for face confidence learning in spherical space. Mathematically, we extend the von Mises Fisher density to its r-radius counterpart and derive a new optimization objective in closed form. Theoretically, the proposed probabilistic framework provably allows for better interpretability, leading to principled feature comparison and pooling. Extensive experimental results on multiple challenging benchmarks confirm our hypothesis and theory, and showcase the advantages of our framework over prior probabilistic methods and spherical deterministic embeddings in various face recognition tasks.

A useful application of event sensing is visual odometry, especially in settings that require high-temporal resolution. The state-of-the-art method of contrast maximisation recovers the motion from a batch of events by maximising the contrast of the image of warped events. However, the cost scales with image resolution and the temporal resolution can be limited by the need for large batch sizes to yield sufficient structure in the contrast image (see supplementary material for demonstration program). In this work, we propose spatiotemporal registration as a compelling technique for event-based rotational motion estimation. We theoretically justify the approach and establish its fundamental and practical advantages over contrast maximisation. In particular, spatiotemporal registration also produces feature tracks as a by-product, which directly supports an efficient visual odometry pipeline with graph-based optimisation for motion averaging. The simplicity of our visual odometry pipeline allows it to process more than 1 M events/second. We also contribute a new event dataset for visual odometry, where motion sequences with large velocity variations were acquired using a high-precision robot arm. Our dataset will be published after the reviewing period.

We present a self-supervised Contrastive Video Representation Learning (CVRL) method to learn spatiotemporal visual representations from unlabeled videos. Our representations are learned using a contrastive loss, where two augmented clips from the same short video are pulled together in the embedding space, while clips from different videos are pushed away. We study what makes for good data augmentations for video self-supervised learning and find that both spatial and temporal information are crucial. We carefully design data augmentations involving spatial and temporal cues. Concretely, we propose a temporally consistent spatial augmentation method to impose strong spatial augmentations on each frame of the video while maintaining the temporal consistency across frames. We also propose a sampling-based temporal augmentation method to avoid overly enforcing invariance on clips that are distant in time. On Kinetics-600, a linear classifier trained on the representations learned by CVRL achieves 70.4% top-1 accuracy with a 3D-ResNet-50 (R3D-50) backbone, outperforming ImageNet supervised pre-training by 15.7% and SimCLR unsupervised pre-training by 18.8% using the same inflated R3D-50. The performance of CVRL can be further improved to 72.9% with a larger R3D-152 (2x filters) backbone, significantly closing the gap between unsupervised and supervised video representation learning. Our code and models will be available at https://github.com/tensorflow/models/tree/master/official/.