CVPR 2020 Papers

We propose a general framework for self-supervised learning of transferable visual representations based on Video-Induced Visual Invariances (VIVI). We consider the implicit hierarchy present in the videos and make use of (i) frame-level invariances (e.g. stability to color and contrast perturbations), (ii) shot/clip-level invariances (e.g. robustness to changes in object orientation and lighting conditions), and (iii) video-level invariances (semantic relationships of scenes across shots/clips), to define a holistic self-supervised loss. Training models using different variants of the proposed framework on videos from the YouTube-8M (YT8M) data set, we obtain state-of-the-art self-supervised transfer learning results on the 19 diverse downstream tasks of the Visual Task Adaptation Benchmark (VTAB), using only 1000 labels per task. We then show how to co-train our models jointly with labeled images, outperforming an ImageNet-pretrained ResNet-50 by 0.8 points with 10x fewer labeled images, as well as the previous best supervised model by 3.7 points using the full ImageNet data set.

The goal of self-supervised learning from images is to construct image representations that are semantically meaningful via pretext tasks that do not require semantic annotations. Many pretext tasks lead to representations that are covariant with image transformations. We argue that, instead, semantic representations ought to be invariant under such transformations. Specifically, we develop Pretext-Invariant Representation Learning (PIRL, pronounced as `pearl') that learns invariant representations based on pretext tasks. We use PIRL with a commonly used pretext task that involves solving jigsaw puzzles. We find that PIRL substantially improves the semantic quality of the learned image representations. Our approach sets a new state-of-the-art in self-supervised learning from images on several popular benchmarks for self-supervised learning. Despite being unsupervised, PIRL outperforms supervised pre-training in learning image representations for object detection. Altogether, our results demonstrate the potential of self-supervised representations with good invariance properties.

We propose a new method for recognizing the pose of objects from a single image that for learning uses only unlabelled videos and a weak empirical prior on the object poses. Video frames differ primarily in the pose of the objects they contain, so our method distils the pose information by analyzing the differences between frames. The distillation uses a new dual representation of the geometry of objects as a set of 2D keypoints, and as a pictorial representation, i.e. a skeleton image. This has three benefits: (1) it provides a tight 'geometric bottleneck' which disentangles pose from appearance, (2) it can leverage powerful image-to-image translation networks to map between photometry and geometry, and (3) it allows to incorporate empirical pose priors in the learning process. The pose priors are obtained from unpaired data, such as from a different dataset or modality such as mocap, such that no annotated image is ever used in learning the pose recognition network. In standard benchmarks for pose recognition for humans and faces, our method achieves state-of-the-art performance among methods that do not require any labelled images for training. Project page: http://www.robots.ox.ac.uk/ vgg/research/unsupervised_pose/

Previous methods on estimating detailed human depth often require supervised training with 'ground truth' depth data. This paper presents a self-supervised method that can be trained on YouTube videos without known depth, which makes training data collection simple and improves the generalization of the learned network. The self-supervised learning is achieved by minimizing a photo-consistency loss, which is evaluated between a video frame and its neighboring frames warped according to the estimated depth and the 3D non-rigid motion of the human body. To solve this non-rigid motion, we first estimate a rough SMPL model at each video frame and compute the non-rigid body motion accordingly, which enables self-supervised learning on estimating the shape details. Experiments demonstrate that our method enjoys better generalization, and performs much better on data in the wild.

Image-level weakly supervised semantic segmentation is a challenging problem that has been deeply studied in recent years. Most of advanced solutions exploit class activation map (CAM). However, CAMs can hardly serve as the object mask due to the gap between full and weak supervisions. In this paper, we propose a self-supervised equivariant attention mechanism (SEAM) to discover additional supervision and narrow the gap. Our method is based on the observation that equivariance is an implicit constraint in fully supervised semantic segmentation, whose pixel-level labels take the same spatial transformation as the input images during data augmentation. However, this constraint is lost on the CAMs trained by image-level supervision. Therefore, we propose consistency regularization on predicted CAMs from various transformed images to provide self-supervision for network learning. Moreover, we propose a pixel correlation module (PCM), which exploits context appearance information and refines the prediction of current pixel by its similar neighbors, leading to further improvement on CAMs consistency. Extensive experiments on PASCAL VOC 2012 dataset demonstrate our method outperforms state-of-the-art methods using the same level of supervision. The code is released online.

Generalized Zero-Shot Learning (GZSL) aims at recognizing both seen and unseen classes by constructing correspondence between visual and semantic embedding. However, existing methods have severely suffered from the strong bias problem, where unseen instances in target domain tend to be recognized as seen classes in source domain. To address this issue, we propose an end-to-end Self-supervised Domain-aware Generative Network (SDGN) by integrating self-supervised learning into feature generating model for unbiased GZSL. The proposed SDGN model enjoys several merits. First, we design a cross-domain feature generating module to synthesize samples with high fidelity based on class embeddings, which involves a novel target domain discriminator to preserve the domain consistency. Second, we propose a self-supervised learning module to investigate inter-domain relationships, where a set of anchors are introduced as a bridge between seen and unseen categories. In the shared space, we pull the distribution of target domain away from source domain, and obtain domain-aware features with high discriminative power for both seen and unseen classes. To our best knowledge, this is the first work to introduce self-supervised learning into GZSL as a learning guidance. Extensive experimental results on five standard benchmarks demonstrate that our model performs favorably against state-of-the-art GZSL methods.

Self-supervised VO methods have shown great success in jointly estimating camera pose and depth from videos. However, like most data-driven methods, existing VO networks suffer from a notable decrease in performance when confronted with scenes different from the training data, which makes them unsuitable for practical applications. In this paper, we propose an online meta-learning algorithm to enable VO networks to continuously adapt to new environments in a self-supervised manner. The proposed method utilizes convolutional long short-term memory (convLSTM) to aggregate rich spatial-temporal information in the past. The network is able to memorize and learn from its past experience for better estimation and fast adaptation to the current frame. When running VO in the open world, in order to deal with the changing environment, we propose an online feature alignment method by aligning feature distributions at different time. Our VO network is able to seamlessly adapt to different environments. Extensive experiments on unseen outdoor scenes, virtual to real world and outdoor to indoor environments demonstrate that our method consistently outperforms state-of-the-art self-supervised VO baselines considerably.

Camera captured human pose is an outcome of several sources of variation. Performance of supervised 3D pose estimation approaches comes at the cost of dispensing with variations, such as shape and appearance, that may be useful for solving other related tasks. As a result, the learned model not only inculcates task-bias but also dataset-bias because of its strong reliance on the annotated samples, which also holds true for weakly-supervised models. Acknowledging this, we propose a self-supervised learning framework to disentangle such variations from unlabeled video frames. We leverage the prior knowledge on human skeleton and poses in the form of a single part-based 2D puppet model, human pose articulation constraints, and a set of unpaired 3D poses. Our differentiable formalization, bridging the representation gap between the 3D pose and spatial part maps, not only facilitates discovery of interpretable pose disentanglement, but also allows us to operate on videos with diverse camera movements. Qualitative results on unseen in-the-wild datasets establish our superior generalization across multiple tasks beyond the primary tasks of 3D pose estimation and part segmentation. Furthermore, we demonstrate state-of-the-art weakly-supervised 3D pose estimation performance on both Human3.6M and MPI-INF-3DHP datasets.

In this paper, we look into the problem of 3D adversary attack, and propose to leverage the internal properties of the point clouds and the adversarial examples to design a new self-robust deep neural network (DNN) based 3D recognition systems. As a matter of fact, on one hand, point clouds are highly structured. Hence for each local part of clean point clouds, it is possible to learn what is it ("part of a bottle") and its relative position ("upper part of a bottle") to the global object center. On the other hand, with the visual quality constraint, 3D adversarial samples often only produce small local perturbations, thus they will roughly keep the original global center but may cause incorrect local relative position estimation. Motivated by these two properties, we use relative position (dubbed as "gather-vector") as the adversarial indicator and propose a new robust gather module. Equipped with this module, we further propose a new self-robust 3D point recognition network. Through extensive experiments, we demonstrate that the proposed method can improve the robustness of the target attack under the white-box setting significantly. For I-FGSM based attack, our method reduces the attack success rate from 94.37 % to 75.69 %. For C&W based attack, our method reduces the attack success rate more than 40.00 %. Moreover, our method is complementary to other types of defense methods to achieve better defense results.

In this paper, we solve the sample shortage problem in the human parsing task. We begin with the self-learning strategy, which generates pseudo-labels for unlabeled data to retrain the model. However, directly using noisy pseudo-labels will cause error amplification and accumulation. Considering the topology structure of human body, we propose a trainable graph reasoning method that establishes internal structural connections between graph nodes to correct two typical errors in the pseudo-labels, i.e., the global structural error and the local consistency error. For the global error, we first transform category-wise features into a high-level graph model with coarse-grained structural information, and then decouple the high-level graph to reconstruct the category features. The reconstructed features have a stronger ability to represent the topology structure of the human body. Enlarging the receptive field of features can effectively reducing the local error. We first project feature pixels into a local graph model to capture pixel-wise relations in a hierarchical graph manner, then reverse the relation information back to the pixels. With the global structural and local consistency modules, these errors are rectified and confident pseudo-labels are generated for retraining. Extensive experiments on the LIP and the ATR datasets demonstrate the effectiveness of our global and local rectification modules. Our method outperforms other state-of-the-art methods in supervised human parsing tasks.

In this paper, we address the problem of rain streaks removal in video by developing a self-learned rain streak removal method, which does not require any clean groundtruth images in the training process. The method is inspired by fact that the adjacent frames are highly correlated and can be regarded as different versions of identical scene, and rain streaks are randomly distributed along the temporal dimension. With this in mind, we construct a two-stage Self-Learned Deraining Network (SLDNet) to remove rain streaks based on both temporal correlation and consistency. In the first stage, SLDNet utilizes the temporal correlations and learns to predict the clean version of the current frame based on its adjacent rain video frames. In the second stage, SLDNet enforces the temporal consistency among different frames. It takes both the current rain frame and adjacent rain video frames to recover structural details. The first stage is responsible for reconstructing main structures, and the second stage is responsible for extracting structural details. We build our network architecture with two sub-tasks, i.e. motion estimation, and rain region detection, and optimize them jointly. Our extensive experiments demonstrate the effectiveness of our method, offering better results both quantitatively and qualitatively.

One crucial aspect of partial domain adaptation (PDA) is how to select the relevant source samples in the shared classes for knowledge transfer. Previous PDA methods tackle this problem by re-weighting the source samples based on their high-level information (deep features). However, since the domain shift between source and target domains, only using the deep features for sample selection is defective. We argue that it is more reasonable to additionally exploit the pixel-level information for PDA problem, as the appearance difference between outlier source classes and target classes is significantly large. In this paper, we propose a reinforced transfer network (RTNet), which utilizes both high-level and pixel-level information for PDA problem. Our RTNet is composed of a reinforced data selector (RDS) based on reinforcement learning (RL), which filters out the outlier source samples, and a domain adaptation model which minimizes the domain discrepancy in the shared label space. Specifically, in the RDS, we design a novel reward based on the reconstruct errors of selected source samples on the target generator, which introduces the pixel-level information to guide the learning of RDS. Besides, we develope a state containing high-level information, which used by the RDS for sample selection. The proposed RDS is a general module, which can be easily integrated into existing DA models to make them fit the PDA situation. Extensive experiments indicate that RTNet can achieve state-of-the-art performance for PDA tasks on several benchmark datasets.

Depth data containing a preponderance of discriminative power in location have been proven beneficial for accurate saliency prediction. However, RGB-D saliency detection methods are also negatively influenced by randomly distributed erroneous or missing regions on the depth map or along the object boundaries. This offers the possibility of achieving more effective inference by well designed models. In this paper, we propose a new framework for accurate RGB-D saliency detection taking account of local and global complementarities from two modalities. This is achieved by designing a complimentary interaction model discriminative enough to simultaneously select useful representation from RGB and depth data, and meanwhile to refine the object boundaries. Moreover, we proposed a compensation-aware loss to further process the information not being considered in the complimentary interaction model, leading to improvement of the generalization ability for challenging scenes. Experiments on six public datasets show that our method outperforms18state-of-the-art methods.

Finding a small subset of data whose linear combination spans other data points, also called column subset selection problem (CSSP), is an important open problem in computer science with many applications in computer vision and deep learning. There are some studies that solve CSSP in a polynomial time complexity w.r.t. the size of the original dataset. A simple and efficient selection algorithm with a linear complexity order, referred to as spectrum pursuit (SP), is proposed that pursuits spectral components of the dataset using available sample points. The proposed non-greedy algorithm aims to iteratively find K data samples whose span is close to that of the first K spectral components of entire data. SP has no parameter to be fine tuned and this desirable property makes it problem-independent. The simplicity of SP enables us to extend the underlying linear model to more complex models such as nonlinear manifolds and graph-based models. The nonlinear extension of SP is introduced as kernel-SP (KSP). The superiority of the proposed algorithms is demonstrated in a wide range of applications.

Fuzzy clustering is known to perform well in real-world applications. Inspired by this observation, we incorporate a fuzzy mechanism into discrete convolutional kernels for 3D point clouds as our first major contribution. The proposed fuzzy kernel is defined over a spherical volume that uses discrete bins. Discrete volumetric division can normally make a kernel vulnerable to boundary effects during learning as well as point density during inference. However, the proposed kernel remains robust to boundary conditions and point density due to the fuzzy mechanism. Our second major contribution comes as the proposal of an efficient graph convolutional network, SegGCN for segmenting point clouds. The proposed network exploits ResNet like blocks in the encoder and 1 x 1 convolutions in the decoder. SegGCN capitalizes on the separable convolution operation of the proposed fuzzy kernel for efficiency. We establish the effectiveness of the SegGCN with the proposed kernel on the challenging S3DIS and ScanNet real-world datasets. Our experiments demonstrate that the proposed network can segment over one million points per second with highly competitive performance.

The majority of current object detectors lack context: class predictions are made independently from other detections. We propose to incorporate context in object detection by post-processing the output of an arbitrary detector to rescore the confidences of its detections. Rescoring is done by conditioning on contextual information from the entire set of detections: their confidences, predicted classes, and positions. We show that AP can be improved by simply reassigning the detection confidence values such that true positives that survive longer (i.e., those with the correct class and large IoU) are scored higher than false positives or detections with small IoU. In this setting, we use a bidirectional RNN with attention for contextual rescoring and introduce a training target that uses the IoU with ground truth to maximize AP for the given set of detections. The fact that our approach does not require access to visual features makes it computationally inexpensive and agnostic to the detection architecture. In spite of this simplicity, our model consistently improves AP over strong pre-trained baselines (Cascade R-CNN and Faster R-CNN with several backbones), particularly by reducing the confidence of duplicate detections (a learned form of non-maximum suppression) and removing out-of-context objects by conditioning on the confidences, classes, positions, and sizes of the co-occurrent detections. Code is available at https://github.com/LourencoVazPato/seeing-without-looking/

We address the dual problems of novel view synthesis and environment reconstruction from hand-held RGBD sensors. Our contributions include 1) modeling highly specular objects, 2) modeling inter-reflections and Fresnel effects, and 3) enabling surface light field reconstruction with the same input needed to reconstruct shape alone. In cases where scene surface has a strong mirror-like material component, we generate highly detailed environment images, revealing room composition, objects, people, buildings, and trees visible through windows. Our approach yields state of the art view synthesis techniques, operates on low dynamic range imagery, and is robust to geometric and calibration errors.

The fusion of multimodal sensor streams, such as camera, lidar, and radar measurements, plays a critical role in object detection for autonomous vehicles, which base their decision making on these inputs. While existing methods exploit redundant information in good environmental conditions, they fail in adverse weather where the sensory streams can be asymmetrically distorted. These rare "edge-case" scenarios are not represented in available datasets, and existing fusion architectures are not designed to handle them. To address this challenge we present a novel multimodal dataset acquired in over 10,000 km of driving in northern Europe. Although this dataset is the first large multimodal dataset in adverse weather, with 100k labels for lidar, camera, radar, and gated NIR sensors, it does not facilitate training as extreme weather is rare. To this end, we present a deep fusion network for robust fusion without a large corpus of labeled training data covering all asymmetric distortions. Departing from proposal-level fusion, we propose a single-shot model that adaptively fuses features, driven by measurement entropy. We validate the proposed method, trained on clean data, on our extensive validation dataset. Code and data are available here https://github.com/princeton-computational-imaging/SeeingThroughFog.

Conventional sensor systems record information about directly visible objects, whereas occluded scene components are considered lost in the measurement process. Non-line-of-sight (NLOS) methods try to recover such hidden objects from their indirect reflections - faint signal components, traditionally treated as measurement noise. Existing NLOS approaches struggle to record these low-signal components outside the lab, and do not scale to large-scale outdoor scenes and high-speed motion, typical in automotive scenarios. In particular, optical NLOS capture is fundamentally limited by the quartic intensity falloff of diffuse indirect reflections. In this work, we depart from visible-wavelength approaches and demonstrate detection, classification, and tracking of hidden objects in large-scale dynamic environments using Doppler radars that can be manufactured at low-cost in series production. To untangle noisy indirect and direct reflections, we learn from temporal sequences of Doppler velocity and position measurements, which we fuse in a joint NLOS detection and tracking network over time. We validate the approach on in-the-wild automotive scenes, including sequences of parked cars or house facades as relay surfaces, and demonstrate low-cost, real-time NLOS in dynamic automotive environments.