CVPR 2021 Papers

In this paper we introduce a Transformer-based approach to video object segmentation (VOS). To address compounding error and scalability issues of prior work, we propose a scalable, end-to-end method for VOS called Sparse Spatiotemporal Transformers (SST). SST extracts per-pixel representations for each object in a video using sparse attention over spatiotemporal features. Our attention-based formulation for VOS allows a model to learn to attend over a history of multiple frames and provides suitable inductive bias for performing correspondence-like computations necessary for solving motion segmentation. We demonstrate the effectiveness of attention-based over recurrent networks in the spatiotemporal domain. Our method achieves competitive results on YouTube-VOS and DAVIS 2017 with improved scalability and robustness to occlusions compared with the state of the art. Code is available at https://github.com/dukebw/SSTVOS.

We introduce an interactive Soft Shadow Network (SSN) to generates controllable soft shadows for image compositing. SSN takes a 2D object mask as input and thus is agnostic to image types such as painting and vector art. An environment light map is used to control the shadow's characteristics, such as angle and softness. SSN employs an Ambient Occlusion Prediction module to predict an intermediate ambient occlusion map, which can be further refined by the user to provides geometric cues to modulate the shadow generation. To train our model, we design an efficient pipeline to produce diverse soft shadow training data using 3D object models. In addition, we propose an inverse shadow map representation to improve model training. We demonstrate that our model produces realistic soft shadows in real-time. Our user studies show that the generated shadows are often indistinguishable from shadows calculated by a physics-based renderer and users can easily use SSN through an interactive application to generate specific shadow effects in minutes.

Recent years have seen flourishing research on both semi-supervised learning and 3D room layout reconstruction. In this work, we explore the intersection of these two fields to advance the research objective of enabling more accurate 3D indoor scene modeling with less labeled data. We propose the first approach to learn representations of room corners and boundaries by using a combination of labeled and unlabeled data for improved layout estimation in a 360-degree panoramic scene. Through extensive comparative experiments, we demonstrate that our approach can advance layout estimation of complex indoor scenes using as few as 20 labeled examples. When coupled with a layout predictor pre-trained on synthetic data, our semi-supervised method matches the fully supervised counterpart using only 12% of the labels. Our work takes an important first step towards robust semi-supervised layout estimation that can enable many applications in 3D perception with limited labeled data.

Self-attention has been successfully applied to video representation learning due to the effectiveness of modeling long range dependencies. Existing approaches build the dependencies merely by computing the pairwise correlations along spatial and temporal dimensions simultaneously. However, spatial correlations and temporal correlations represent different contextual information of scenes and temporal reasoning. Intuitively, learning spatial contextual information first will benefit temporal modeling. In this paper, we propose a separable self-attention (SSA) module, which models spatial and temporal correlations sequentially, so that spatial contexts can be efficiently used in temporal modeling. By adding SSA module into 2D CNN, we build a SSA network (SSAN) for video representation learning. On the task of video action recognition, our approach outperforms state-of-the-art methods on Something-Something and Kinetics-400 datasets. Our models often outperform counterparts with shallower network and less modality. We further verify the semantic learning ability of our method in visual-language task of video retrieval, which showcase the homogeneity of video representations and text embeddings. On MSR-VTT and Youcook2 datasets, video representations learnt by SSA significantly improve the state-of-the-art methods.

Deep CNNs have achieved significant successes in image processing and its applications, including single image super-resolution (SR). However, conventional methods still resort to some predetermined integer scaling factors, e.g., x2 or x4. Thus, they are difficult to be applied when arbitrary target resolutions are required. Recent approaches extend the scope to real-valued upsampling factors, even with varying aspect ratios to handle the limitation. In this paper, we propose the SRWarp framework to further generalize the SR tasks toward an arbitrary image transformation. We interpret the traditional image warping task, specifically when the input is enlarged, as a spatially-varying SR problem. We also propose several novel formulations, including the adaptive warping layer and multiscale blending, to reconstruct visually favorable results in the transformation process. Compared with previous methods, we do not constrain the SR model on a regular grid but allow numerous possible deformations for flexible and diverse image editing. Extensive experiments and ablation studies justify the necessity and demonstrate the advantage of the proposed SRWarp method under various transformations.

Geometric characteristic plays an important role in the representation of an object in 3D point clouds. For example, large objects often contain more points, while small ones contain fewer points. The point clouds of objects near the capture device are denser, while those of distant objects are sparser. These issues bring new challenges to 3D object detection, especially under the domain adaptation scenarios. In this work, we propose a new cross-dataset 3D object detection method named Scale-aware and Range-aware Domain Adaptation Network (SRDAN). We take advantage of the geometric characteristics of 3D data (i.e., size and distance), and propose the scale-aware domain alignment and the range-aware domain alignment strategies to guide the distribution alignment between two domains. For scale-aware domain alignment, we design a 3D voxel-based feature pyramid network to extract multi-scale semantic voxel features, and align the features and instances with similar scales between two domains. For range-aware domain alignment, we introduce a range-guided domain alignment module to align the features of objects according to their distance to the capture device. Extensive experiments under three different scenarios demonstrate the effectiveness of our SRDAN approach, and comprehensive ablation study also validates the importance of geometric characteristics for cross-dataset 3D object detection.

We present SPSG, a novel approach to generate high-quality, colored 3D models of scenes from RGB-D scan observations by learning to infer unobserved scene geometry and color in a self-supervised fashion. Our self-supervised approach learns to jointly inpaint geometry and color by correlating an incomplete RGB-D scan with a more complete version of that scan. Notably, rather than relying on 3D reconstruction losses to inform our 3D geometry and color reconstruction, we propose adversarial and perceptual losses operating on 2D renderings in order to achieve high-resolution, high-quality colored reconstructions of scenes. This exploits the high-resolution, self-consistent signal from individual raw RGB-D frames, in contrast to fused 3D reconstructions of the frames which exhibit inconsistencies from view-dependent effects, such as color balancing or pose inconsistencies. Thus, by informing our 3D scene generation directly through 2D signal, we produce high-quality colored reconstructions of 3D scenes, outperforming state of the art on both synthetic and real data.

The ability to navigate like a human towards a language-guided target from anywhere in a 3D embodied environment is one of the 'holy grail' goals of intelligent robots. Most visual navigation benchmarks, however, focus on navigating toward a target from a fixed starting point, guided by an elaborate set of instructions that depicts step-by-step. This approach deviates from real-world problems in which human-only describes what the object and its surrounding look like and asks the robot to start navigation from anywhere. Accordingly, in this paper, we introduce a Scenario Oriented Object Navigation (SOON) task. In this task, an agent is required to navigate from an arbitrary position in a 3D embodied environment to localize a target following a scene description. To give a promising direction to solve this task, we propose a novel graph-based exploration (GBE) method, which models the navigation state as a graph and introduces a novel graph-based exploration approach to learn knowledge from the graph and stabilize training by learning sub-optimal trajectories. We also propose a new large-scale benchmark named From Anywhere to Object (FAO) dataset. To avoid target ambiguity, the descriptions in FAO provide rich semantic scene information includes: object attribute, object relationship, region description, and nearby region description. Our experiments reveal that the proposed GBE outperforms various state-of-the-arts on both FAO and R2R datasets. And the ablation studies on FAO validates the quality of the dataset.

Compared to feature point detection and description, detecting and matching line segments offer additional challenges. Yet, line features represent a promising complement to points for multi-view tasks. Lines are indeed well-defined by the image gradient, frequently appear even in poorly textured areas and offer robust structural cues. We thus hereby introduce the first joint detection and description of line segments in a single deep network. Thanks to a self-supervised training, our method does not require any annotated line labels and can therefore generalize to any dataset. Our detector offers repeatable and accurate localization of line segments in images, departing from the wireframe parsing approach. Leveraging the recent progresses in descriptor learning, our proposed line descriptor is highly discriminative, while remaining robust to viewpoint changes and occlusions. We evaluate our approach against previous line detection and description methods on several multi-view datasets created with homographic warps as well as real-world viewpoint changes. Our full pipeline yields higher repeatability, localization accuracy and matching metrics, and thus represents a first step to bridge the gap with learned feature points methods. Code and trained weights are available at https://github.com/cvg/SOLD2.

We tackle the problem of place recognition from point cloud data and introduce a self-attention and orientation encoding network (SOE-Net) that fully explores the relationship between points and incorporates long-range context into point-wise local descriptors. Local information of each point from eight orientations is captured in a PointOE module, whereas long-range feature dependencies among local descriptors are captured with a self-attention unit. Moreover, we propose a novel loss function called Hard Positive Hard Negative quadruplet loss (HPHN quadruplet), that achieves better performance than the commonly used metric learning loss. Experiments on various benchmark datasets demonstrate superior performance of the proposed network over the current state-of-the-art approaches. Our code is released publicly at https://github.com/Yan-Xia/SOE-Net.

We present SMURF, a method for unsupervised learning of optical flow that improves state of the art on all benchmarks by 36% to 40% and even outperforms several supervised approaches such as PWC-Net and FlowNet2. Our method integrates architecture improvements from supervised optical flow, i.e. the RAFT model, with new ideas for unsupervised learning that include a novel unsupervised sequence loss and self-supervision loss, a technique for handling out-of-frame motion, and an approach for learning effectively from multi-frame video data while still only requiring two frames for inference.

In this paper we introduce SMPLicit, a novel generative model to jointly represent body pose, shape and clothing geometry. In contrast to existing learning-based approaches that require training specific models for each type of garment, SMPLicit can represent in a unified manner different garment topologies (e.g. from sleeveless tops to hoodies and to open jackets), while controlling other properties like the garment size or tightness/looseness. We show our model to be applicable to a large variety of garments including T-shirts, hoodies, jackets, shorts, pants, skirts, shoes and even hair. The representation flexibility of SMPLicit builds upon an implicit model conditioned with the SMPL human body parameters and a learnable latent space which is semantically interpretable and aligned with the clothing attributes. The proposed model is fully differentiable, allowing for its use into larger end-to-end trainable systems. In the experimental section, we demonstrate SMPLicit can be readily used for fitting 3D scans and for 3D reconstruction in images of dressed people. In both cases we are able to go beyond state of the art, by retrieving complex garment geometries, handling situations with multiple clothing layers and providing a tool for easy outfit editing. To stimulate further research in this direction, we will make our code and model publicly available at https://link/smplicit/.

Despite stereo matching accuracy has greatly improved by deep learning in the last few years, recovering sharp boundaries and high-resolution outputs efficiently remains challenging. In this paper, we propose Stereo Mixture Density Networks (SMD-Nets), a simple yet effective learning framework compatible with a wide class of 2D and 3D architectures which ameliorates both issues. Specifically, we exploit bimodal mixture densities as output representation and show that this allows for sharp and precise disparity estimates near discontinuities while explicitly modeling the aleatoric uncertainty inherent in the observations. Moreover, we formulate disparity estimation as a continuous problem in the image domain, allowing our model to query disparities at arbitrary spatial precision. We carry out comprehensive experiments on a new high-resolution and highly realistic synthetic stereo dataset, consisting of stereo pairs at 8Mpx resolution, as well as on real-world stereo datasets. Our experiments demonstrate increased depth accuracy near object boundaries and prediction of ultra high-resolution disparity maps on standard GPUs. We demonstrate the flexibility of our technique by improving the performance of a variety of stereo backbones.

Most existing distance metric learning approaches use fully labeled data to learn the sample similarities in an embedding space. We present a self-training framework, SLADE, to improve retrieval performance by leveraging additional unlabeled data. We first train a teacher model on the labeled data and use it to generate pseudo labels for the unlabeled data. We then train a student model on both labels and pseudo labels to generate final feature embeddings. We use self-supervised representation learning to initialize the teacher model. To better deal with noisy pseudo labels generated by the teacher network, we design a new feature basis learning component for the student network, which learns basis functions of feature representations for unlabeled data. The learned basis vectors better measure the pairwise similarity and are used to select high-confident samples for training the student network. We evaluate our method on standard retrieval benchmarks: CUB-200, Cars-196 and In-shop. Experimental results demonstrate that with additional unlabeled data, our approach significantly improves the performance over the state-of-the-art methods.

The bottleneck of computation burden limits the widespread use of the 2nd order optimization algorithms for training deep neural networks. In this paper, we present a computationally efficient approximation for natural gradient descent, named Swift Kronecker-Factored Approximate Curvature (SKFAC), which combines Kronecker factorization and a fast low-rank matrix inversion technique. Our research aims at both fully connected and convolutional layers. For the fully connected layers, by utilizing the low-rank property of Kronecker factors of Fisher information matrix, our method only requires inverting a small matrix to approximate the curvature with desirable accuracy. For convolutional layers, we propose a way with two strategies to save computational efforts without affecting the empirical performance by reducing across the spatial dimension or receptive fields of feature maps. Specifically, we propose two effective dimension reduction methods for this purpose: Spatial Subsampling and Reduce Sum. Experimental results of training several deep neural networks on Cifar-10 and ImageNet-1k datasets demonstrate that SKFAC can capture the main curvature and yield comparative performance to K-FAC. The proposed method bridges the wall-clock time gap between the 1st and 2nd order algorithms.

Pan-sharpening is a process of merging a high-resolution (HR) panchromatic (PAN) image and its corresponding low-resolution (LR) multi-spectral (MS) image to create an HR-MS and pan-sharpened image. However, due to the different sensors' locations, characteristics and acquisition time, PAN and MS image pairs often tend to have various amounts of misalignment. Conventional deep-learning-based methods that were trained with such misaligned PAN-MS image pairs suffer from diverse artifacts such as double-edge and blur artifacts in the resultant PAN-sharpened images. In this paper, we propose a novel framework called shift-invariant pan-sharpening with moving object alignment (SIPSA-Net) which is the first method to take into account such large misalignment of moving object regions for PAN sharpening. The SISPA-Net has a feature alignment module (FAM) that can adjust one feature to be aligned to another feature, even between the two different PAN and MS domains. For better alignment in pan-sharpened images, a shift-invariant spectral loss is newly designed, which ignores the inherent misalignment in the original MS input, thereby having the same effect as optimizing the spectral loss with a well-aligned MS image. Extensive experimental results show that our SIPSA-Net can generate pan-sharpened images with remarkable improvements in terms of visual quality and alignment, compared to the state-of-the-art methods.

Pedestrian trajectory prediction is a key technology in autopilot, which remains to be very challenging due to complex interactions between pedestrians. However, previous works based on dense undirected interaction suffer from modeling superfluous interactions and neglect of trajectory motion tendency, and thus inevitably result in a considerable deviance from the reality. To cope with these issues, we present a Sparse Graph Convolution Network (SGCN) for pedestrian trajectory prediction. Specifically, the SGCN explicitly models the sparse directed interaction with a sparse directed spatial graph to capture adaptive interaction pedestrians. Meanwhile, we use a sparse directed temporal graph to model the motion tendency, thus to facilitate the prediction based on the observed direction. Finally, parameters of a bi-Gaussian distribution for trajectory prediction are estimated by fusing the above two sparse graphs. We evaluate our proposed method on the ETH and UCY datasets, and the experimental results show our method outperforms comparative state-of-the-art methods by 9% in Average Displacement Error (ADE) and 13% in Final Displacement Error (FDE). Notably, visualizations indicate that our method can capture adaptive interactions between pedestrians and their effective motion tendencies.

Video instance segmentation (VIS) is a new and critical task in computer vision. To date, top-performing VIS methods extend the two-stage Mask R-CNN by adding a tracking branch, leaving plenty of room for improvement. In contrast, we approach the VIS task from a new perspective and propose a one-stage spatial granularity network (SG-Net). SG-Net demonstrates four advantages: 1) Our task heads (detection, segmentation, and tracking) are crafted interdependently so they can effectively share features and enjoy the joint optimization; 2) Each of our task predictions avoids using proposal-based RoI features, resulting in much reduced runtime complexity per instance; 3) Our mask prediction is dynamically performed on the sub-regions of each detected instance, leading to high-quality masks of fine granularity; 4) Our tracking head models objects' centerness movements for tracking, which effectively enhances the tracking robustness to different object appearances. In evaluation, we present state-of-the-art comparisons on the YouTube-VIS dataset. Extensive experiments demonstrate that our compact one-stage method can achieve improved performance in both accuracy and inference speed. We hope our SG-Net could serve as a simple yet strong baseline for the VIS task. Code will be available.

We present Self-Ensembling Single-Stage object Detector (SE-SSD) for accurate and efficient 3D object detection in outdoor point clouds. Our key focus is on exploiting both soft and hard targets with our formulated constraints to jointly optimize the model, without introducing extra computation in the inference. Specifically, SE-SSD contains a pair of teacher and student SSDs, in which we design an effective IoU-based matching strategy to filter soft targets from the teacher and formulate a consistency loss to align student predictions with them. Also, to maximize the distilled knowledge for ensembling the teacher, we design a new augmentation scheme to produce shape-aware augmented samples to train the student, aiming to encourage it to infer complete object shapes. Lastly, to better exploit hard targets, we design an ODIoU loss to supervise the student with constraints on the predicted box centers and orientations. Our SE-SSD attains top performance compared with all prior published works. Also, it attains top precisions for car detection in the KITTI benchmark (ranked 1st and 2nd on the BEV and 3D leaderboards, respectively) with an ultra-high inference speed. The code is available at https://github.com/Vegeta2020/SE-SSD.