CVPR 2020 Papers

Robust detection and tracking of objects is crucial for the deployment of autonomous vehicle technology. Image based benchmark datasets have driven development in computer vision tasks such as object detection, tracking and segmentation of agents in the environment. Most autonomous vehicles, however, carry a combination of cameras and range sensors such as lidar and radar. As machine learning based methods for detection and tracking become more prevalent, there is a need to train and evaluate such methods on datasets containing range sensor data along with images. In this work we present nuTonomy scenes (nuScenes), the first dataset to carry the full autonomous vehicle sensor suite: 6 cameras, 5 radars and 1 lidar, all with full 360 degree field of view. nuScenes comprises 1000 scenes, each 20s long and fully annotated with 3D bounding boxes for 23 classes and 8 attributes. It has 7x as many annotations and 100x as many images as the pioneering KITTI dataset. We define novel 3D detection and tracking metrics. We also provide careful dataset analysis as well as baselines for lidar and image based detection and tracking. Data, development kit and more information are available online.

Humans can continuously learn new knowledge as their experience grows. In contrast, previous learning in deep neural networks can quickly fade out when they are trained on a new task. In this paper, we hypothesize this problem can be avoided by learning a set of generalized parameters, that are neither specific to old nor new tasks. In this pursuit, we introduce a novel meta-learning approach that seeks to maintain an equilibrium between all the encountered tasks. This is ensured by a new meta-update rule which avoids catastrophic forgetting. In comparison to previous meta-learning techniques, our approach is task-agnostic. When presented with a continuum of data, our model automatically identifies the task and quickly adapts to it with just a single update. We perform extensive experiments on five datasets in a class-incremental setting, leading to significant improvements over the state of the art methods (e.g., a 21.3% boost on CIFAR100 with 10 incremental tasks). Specifically, on large-scale datasets that generally prove difficult cases for incremental learning, our approach delivers absolute gains as high as 19.1% and 7.4% on ImageNet and MS-Celeb datasets, respectively.

Many real-world applications in augmented reality (AR), 3D mapping, and robotics require both fast and accurate estimation of camera poses and scales from multiple images captured by multiple cameras or a single moving camera. Achieving high speed and maintaining high accuracy in a pose-and-scale estimator are often conflicting goals. To simultaneously achieve both, we exploit a priori knowledge about the solution space. We present gDLS*, a generalized-camera-model pose-and-scale estimator that utilizes rotation and scale priors. gDLS* allows an application to flexibly weigh the contribution of each prior, which is important since priors often come from noisy sensors. Compared to state-of-the-art generalized-pose-and-scale estimators (e.g., gDLS), our experiments on both synthetic and real data consistently demonstrate that gDLS* accelerates the estimation process and improves scale and pose accuracy.

In this paper, we explore the space-time video super-resolution task, which aims to generate a high-resolution (HR) slow-motion video from a low frame rate (LFR), low-resolution (LR) video. A simple solution is to split it into two sub-tasks: video frame interpolation (VFI) and video super-resolution (VSR). However, temporal interpolation and spatial super-resolution are intra-related in this task. Two-stage methods cannot fully take advantage of the natural property. In addition, state-of-the-art VFI or VSR networks require a large frame-synthesis or reconstruction module for predicting high-quality video frames, which makes the two-stage methods have large model sizes and thus be time-consuming. To overcome the problems, we propose a one-stage space-time video super-resolution framework, which directly synthesizes an HR slow-motion video from an LFR, LR video. Rather than synthesizing missing LR video frames as VFI networks do, we firstly temporally interpolate LR frame features in missing LR video frames capturing local temporal contexts by the proposed feature temporal interpolation network. Then, we propose a deformable ConvLSTM to align and aggregate temporal information simultaneously for better leveraging global temporal contexts. Finally, a deep reconstruction network is adopted to predict HR slow-motion video frames. Extensive experiments on benchmark datasets demonstrate that the proposed method not only achieves better quantitative and qualitative performance but also is more than three times faster than recent two-stage state-of-the-art methods, e.g., DAIN+EDVR and DAIN+RBPN.

Quantization is a promising approach for reducing the inference time and memory footprint of neural networks. However, most existing quantization methods require access to the original training dataset for retraining during quantization. This is often not possible for applications with sensitive or proprietary data, e.g., due to privacy and security concerns. Existing zero-shot quantization methods use different heuristics to address this, but they result in poor performance, especially when quantizing to ultra-low precision. Here, we propose \OURS, a novel zero-shot quantization framework to address this. \OURS enables mixed-precision quantization without any access to the training or validation data. This is achieved by optimizing for a Distilled Dataset, which is engineered to match the statistics of batch normalization across different layers of the network. \OURS supports both uniform and mixed-precision quantization. For the latter, we introduce a novel Pareto frontier based method to automatically determine the mixed-precision bit setting for all layers, with no manual search involved. We extensively test our proposed method on a diverse set of models, including ResNet18/50/152, MobileNetV2, ShuffleNet, SqueezeNext, and InceptionV3 on ImageNet, as well as RetinaNet-ResNet50 on the Microsoft COCO dataset. In particular, we show that \OURS can achieve 1.71% higher accuracy on MobileNetV2, as compared to the recently proposed DFQ [??] method. Importantly, \OURS has a very low computational overhead, and it can finish the entire quantization process in less than 30s (0.5% of one epoch training time of ResNet50 on ImageNet). We have open-sourced the \OURS framework(https://github.com/amirgholami/ZeroQ).

The paper presents a novel method, Zero-Reference Deep Curve Estimation (Zero-DCE), which formulates light enhancement as a task of image-specific curve estimation with a deep network. Our method trains a lightweight deep network, DCE-Net, to estimate pixel-wise and high-order curves for dynamic range adjustment of a given image. The curve estimation is specially designed, considering pixel value range, monotonicity, and differentiability. Zero-DCE is appealing in its relaxed assumption on reference images, i.e., it does not require any paired or unpaired data during training. This is achieved through a set of carefully formulated non-reference loss functions, which implicitly measure the enhancement quality and drive the learning of the network. Our method is efficient as image enhancement can be achieved by an intuitive and simple nonlinear curve mapping. Despite its simplicity, we show that it generalizes well to diverse lighting conditions. Extensive experiments on various benchmarks demonstrate the advantages of our method over state-of-the-art methods qualitatively and quantitatively. Furthermore, the potential benefits of our Zero-DCE to face detection in the dark are discussed.

Graph matching (GM), as a longstanding problem in computer vision and pattern recognition, still suffers from numerous cluttered outliers in practical applications. To address this issue, we present the zero-assignment constraint (ZAC) for approaching the graph matching problem in the presence of outliers. The underlying idea is to suppress the matchings of outliers by assigning zero-valued vectors to the potential outliers in the obtained optimal correspondence matrix. We provide elaborate theoretical analysis to the problem, i.e., GM with ZAC, and figure out that the GM problem with and without outliers are intrinsically different, which enables us to put forward a sufficient condition to construct valid and reasonable objective function. Consequently, we design an efficient outlier-robust algorithm to significantly reduce the incorrect or redundant matchings caused by numerous outliers. Extensive experiments demonstrate that our method can achieve the state-of-the-art performance in terms of accuracy and efficiency, especially in the presence of numerous outliers.

An integral part of video analysis and surveillance is temporal activity detection, which means to simultaneously recognize and localize activities in long untrimmed videos. Currently, the most effective methods of temporal activity detection are based on deep learning, and they typically perform very well with large scale annotated videos for training. However, these methods are limited in real applications due to the unavailable videos about certain activity classes and the time-consuming data annotation. To solve this challenging problem, we propose a novel task setting called zero-shot temporal activity detection (ZSTAD), where activities that have never been seen in training can still be detected. We design an end-to-end deep network based on R-C3D as the architecture for this solution. The proposed network is optimized with an innovative loss function that considers the embeddings of activity labels and their super-classes while learning the common semantics of seen and unseen activities. Experiments on both the THUMOS'14 and the Charades datasets show promising performance in terms of detecting unseen activities.

We introduce a new local sparse attention layer that preserves two-dimensional geometry and locality. We show that by just replacing the dense attention layer of SAGAN with our construction, we obtain very significant FID, Inception score and pure visual improvements. FID score is improved from 18.65 to 15.94 on ImageNet, keeping all other parameters the same. The sparse attention patterns that we propose for our new layer are designed using a novel information theoretic criterion that uses information flow graphs. We also present a novel way to invert Generative Adversarial Networks with attention. Our method uses the attention layer of the discriminator to create an innovative loss function. This allows us to visualize the newly introduced attention heads and show that they indeed capture interesting aspects of two-dimensional geometry of real images.

The body pose of a person wearing a camera is of great interest for applications in augmented reality, healthcare, and robotics, yet much of the person's body is out of view for a typical wearable camera. We propose a learning-based approach to estimate the camera wearer's 3D body pose from egocentric video sequences. Our key insight is to leverage interactions with another person---whose body pose we can directly observe---as a signal inherently linked to the body pose of the first-person subject. We show that since interactions between individuals often induce a well-ordered series of back-and-forth responses, it is possible to learn a temporal model of the interlinked poses even though one party is largely out of view. We demonstrate our idea on a variety of domains with dyadic interaction and show the substantial impact on egocentric body pose estimation, which improves the state of the art.

This paper presents X3D, a family of efficient video networks that progressively expand a tiny 2D image classification architecture along multiple network axes, in space, time, width and depth. Inspired by feature selection methods in machine learning, a simple stepwise network expansion approach is employed that expands a single axis in each step, such that good accuracy to complexity trade-off is achieved. To expand X3D to a specific target complexity, we perform progressive forward expansion followed by backward contraction. X3D achieves state-of-the-art performance while requiring 4.8x and 5.5x fewer multiply-adds and parameters for similar accuracy as previous work. Our most surprising finding is that networks with high spatiotemporal resolution can perform well, while being extremely light in terms of network width and parameters. We report competitive accuracy at unprecedented efficiency on video classification and detection benchmarks. Code is available at: https://github.com/facebookresearch/SlowFast.

Recent progress on fine-grained visual recognition and visual question answering has featured Bilinear Pooling, which effectively models the 2nd order interactions across multi-modal inputs. Nevertheless, there has not been evidence in support of building such interactions concurrently with attention mechanism for image captioning. In this paper, we introduce a unified attention block --- X-Linear attention block, that fully employs bilinear pooling to selectively capitalize on visual information or perform multi-modal reasoning. Technically, X-Linear attention block simultaneously exploits both the spatial and channel-wise bilinear attention distributions to capture the 2^ nd order interactions between the input single-modal or multi-modal features. Higher and even infinity order feature interactions are readily modeled through stacking multiple X-Linear attention blocks and equipping the block with Exponential Linear Unit (ELU) in a parameter-free fashion, respectively. Furthermore, we present X-Linear Attention Networks (dubbed as X-LAN) that novelly integrates X-Linear attention block(s) into image encoder and sentence decoder of image captioning model to leverage higher order intra- and inter-modal interactions. The experiments on COCO benchmark demonstrate that our X-LAN obtains to-date the best published CIDEr performance of 132.0% on COCO Karpathy test split. When further endowing Transformer with X-Linear attention blocks, CIDEr is boosted up to 132.8%. Source code is available at https://github.com/Panda-Peter/image-captioning.

We present a novel method for inserting objects, specifically humans, into existing images, such that they blend in a photorealistic manner, while respecting the semantic context of the scene. Our method involves three subnetworks: the first generates the semantic map of the new person, given the pose of the other persons in the scene and an optional bounding box specification. The second network renders the pixels of the novel person and its blending mask, based on specifications in the form of multiple appearance components. A third network refines the generated face in order to match those of the target person. Our experiments present convincing high-resolution outputs in this novel and challenging application domain. In addition, the three networks are evaluated individually, demonstrating for example, state of the art results in pose transfer benchmarks.

Camera calibration is an essential first step in setting up 3D Computer Vision systems. Commonly used parametric camera models are limited to a few degrees of freedom and thus often do not optimally fit to complex real lens distortion. In contrast, generic camera models allow for very accurate calibration due to their flexibility. Despite this, they have seen little use in practice. In this paper, we argue that this should change. We propose a calibration pipeline for generic models that is fully automated, easy to use, and can act as a drop-in replacement for parametric calibration, with a focus on accuracy. We compare our results to parametric calibrations. Considering stereo depth estimation and camera pose estimation as examples, we show that the calibration error acts as a bias on the results. We thus argue that in contrast to current common practice, generic models should be preferred over parametric ones whenever possible. To facilitate this, we released our calibration pipeline at https://github.com/puzzlepaint/camera_calibration, making both easy-to-use and accurate camera calibration available to everyone.

With the rapid growth of e-commerce and the popularity of online shopping, fashion retrieval has received considerable attention in the computer vision community. Different from the existing works that mainly focus on identical or similar fashion item retrieval, in this paper, we aim to study the plagiarized clothes retrieval which is somewhat ignored in the academic community while itself has great application value. One of the key challenges is that plagiarized clothes are usually modified in a certain region on the original design to escape the supervision by traditional retrieval methods. To relieve it, we propose a novel network named Plagiarized-Search-Net (PS-Net) based on regional representation, where we utilize the landmarks to guide the learning of regional representations and compare fashion items region by region. Besides, we propose a new dataset named Plagiarized Fashion for plagiarized clothes retrieval, which provides a meaningful complement to the existing fashion retrieval field. Experiments on Plagiarized Fashion dataset verify that our approach is superior to other instance-level counterparts for plagiarized clothes retrieval, showing a promising result for original design protection. Moreover, our PS-Net can also be adapted to traditional fashion retrieval and landmark estimation tasks and achieves the state-of-the-art performance on the DeepFashion and DeepFashion2 datasets.

Pedestrian detection benefits greatly from deep convolutional neural networks (CNNs). However, it is inherently hard for CNNs to handle situations in the presence of occlusion and scale variation. In this paper, we propose W^3Net, which attempts to address above challenges by decomposing the pedestrian detection task into Where, What and Whether problem directing against pedestrian localization, scale prediction and classification correspondingly. Specifically, for a pedestrian instance, we formulate its feature by three steps. i) We generate a bird view map, which is naturally free from occlusion issues, and scan all points on it to look for suitable locations for each pedestrian instance. ii) Instead of utilizing pre-fixed anchors, we model the interdependency between depth and scale aiming at generating depth-guided scales at different locations for better matching instances of different sizes. iii) We learn a latent vector shared by both visual and corpus space, by which false positives with similar vertical structure but lacking human partial features would be filtered out. We achieve state-of-the-art results on widely used datasets (Citypersons and Caltech). In particular. when evaluating on heavy occlusion subset, our results reduce MR^ -2 from 49.3% to 18.7% on Citypersons, and from 45.18% to 28.33% on Caltech.

In this paper, we consider a novel task, Spatio-Temporal Video Grounding for Multi-Form Sentences (STVG). Given an untrimmed video and a declarative/interrogative sentence depicting an object, STVG aims to localize the spatio-temporal tube of the queried object. STVG has two challenging settings: (1) We need to localize spatio-temporal object tubes from untrimmed videos, where the object may only exist in a very small segment of the video; (2) We deal with multi-form sentences, including the declarative sentences with explicit objects and interrogative sentences with unknown objects. Existing methods cannot tackle the STVG task due to the ineffective tube pre-generation and the lack of object relationship modeling. Thus, we then propose a novel Spatio-Temporal Graph Reasoning Network (STGRN) for this task. First, we build a spatio-temporal region graph to capture the region relationships with temporal object dynamics, which involves the implicit and explicit spatial subgraphs in each frame and the temporal dynamic subgraph across frames. We then incorporate textual clues into the graph and develop the multi-step cross-modal graph reasoning. Next, we introduce a spatio-temporal localizer with a dynamic selection method to directly retrieve the spatio-temporal tubes without tube pre-generation. Moreover, we contribute a large-scale video grounding dataset VidSTG based on video relation dataset VidOR. The extensive experiments demonstrate the effectiveness of our method.

Dynamic scene understanding is an essential capability in robotics and VR/AR. In this paper we propose Co-Section, an optimization-based approach to 3D dynamic scene reconstruction, which infers hidden shape information from intersection constraints. An object-level dynamic SLAM frontend detects, segments, tracks and maps dynamic objects in the scene. Our optimization backend completes the shapes using hull and intersection constraints between the objects. In experiments, we demonstrate our approach on real and synthetic dynamic scene datasets. We also assess the shape completion performance of our method quantitatively. To the best of our knowledge, our approach is the first method to incorporate such physical plausibility constraints on object intersections for shape completion of dynamic objects in an energy minimization framework.

Cross-view geo-localization is the problem of estimating the position and orientation (latitude, longitude and azimuth angle) of a camera at ground level given a large-scale database of geo-tagged aerial (eg., satellite) images. Existing approaches treat the task as a pure location estimation problem by learning discriminative feature descriptors, but neglect orientation alignment. It is well-recognized that knowing the orientation between ground and aerial images can significantly reduce matching ambiguity between these two views, especially when the ground-level images have a limited Field of View (FoV) instead of a full field-of-view panorama. Therefore, we design a Dynamic Similarity Matching network to estimate cross-view orientation alignment during localization. In particular, we address the cross-view domain gap by applying a polar transform to the aerial images to approximately align the images up to an unknown azimuth angle. Then, a two-stream convolutional network is used to learn deep features from the ground and polar-transformed aerial images. Finally, we obtain the orientation by computing the correlation between cross-view features, which also provides a more accurate measure of feature similarity, improving location recall. Experiments on standard datasets demonstrate that our method significantly improves state-of-the-art performance. Remarkably, we improve the top-1 location recall rate on the CVUSA dataset by a factor of 1.5x for panoramas with known orientation, by a factor of 3.3x for panoramas with unknown orientation, and by a factor of 6x for 180-degree FoV images with unknown orientation.