CVPR 2021 论文列表

We consider the problem of generating realistic traffic scenes automatically. Existing methods typically insert actors into the scene according to a set of hand-crafted heuristics and are limited in their ability to model the true complexity and diversity of real traffic scenes, thus inducing a content gap between synthesized traffic scenes versus real ones. As a result, existing simulators lack the fidelity necessary to train and test self-driving vehicles. To address this limitation, we present SceneGen, a neural autoregressive model of traffic scenes that eschews the need for rules and heuristics. In particular, given the ego-vehicle state and a high definition map of surrounding area, SceneGen inserts actors of various classes into the scene and synthesizes their sizes, orientations, and velocities. We demonstrate on two large-scale datasets SceneGen's ability to faithfully model distributions of real traffic scenes. Moreover, we show that SceneGen coupled with sensor simulation can be used to train perception models that generalize to the real world.

Humans learn from life events to form intuitions towards the understanding of visual environments and languages. Envision that you are instructed by a high-level instruction, "Go to the bathroom in the master bedroom and replace the blue towel on the left wall", what would you possibly do to carry out the task? Intuitively, we comprehend the semantics of the instruction to form an overview of where a bathroom is and what a blue towel is in mind; then, we navigate to the target location by consistently matching the bathroom appearance in mind with the current scene. In this paper, we present an agent that mimics such human behaviors. Specifically, we focus on the Remote Embodied Visual Referring Expression in Real Indoor Environments task, called REVERIE, where an agent is asked to correctly localize a remote target object specified by a concise high-level natural language instruction, and propose a two-stage training pipeline. In the first stage, we pre-train the agent with two cross-modal alignment sub-tasks, namely the Scene Grounding task and the Object Grounding task. The agent learns where to stop in the Scene Grounding task and what to attend to in the Object Grounding task respectively. Then, to generate action sequences, we propose a memory-augmented attentive action decoder to smoothly fuse the pre-trained vision and language representations with the agent's past memory experiences. Without bells and whistles, experimental results show that our method outperforms previous state-of-the-art(SOTA) significantly, demonstrating the effectiveness of our method.

We revisit human motion synthesis, a task useful in various real-world applications, in this paper. Whereas a number of methods have been developed previously for this task, they are often limited in two aspects: 1) focus on the poses while leaving the location movement behind, and 2) ignore the impact of the environment on the human motion. In this paper, we propose a new framework, with the interaction between the scene and the human motion is taken into account. Considering the uncertainty of human motion, we formulate this task as a generative task, whose objective is to generate plausible human motion conditioned on both the scene and the human's initial position. This framework factorizes the distribution of human motions into a distribution of movement trajectories conditioned on scenes and that of body pose dynamics conditioned on both scenes and trajectories. We further derive a GAN-based learning approach, with discriminators to enforce the compatibility between the human motion and the contextual scene as well as the 3D-to-2D projection constraints. We assess the effectiveness of the proposed method on two challenging datasets, which cover both synthetic and real-world environmentemphasizes local structural constraints via depth-map crops, and a projection discriminator that emphasizes global structural constraints via 3D-to-2D motion projections. The effectiveness of our framework is comprehensively evaluated on two large challenging datasets, covering both a synthetic environment (GTA-IM) and a real environment (PROX)

Image super-resolution, which is often regarded as a preprocessing procedure of scene text recognition, aims to recover the realistic features from a low-resolution text image. It has always been challenging due to large variations in text shapes, fonts, backgrounds, etc. However, most existing methods employ generic super-resolution frameworks to handle scene text images while ignoring text-specific properties such as text-level layouts and character-level details. In this paper, we establish a text-focused super-resolution framework, called Scene Text Telescope (STT). In terms of text-level layouts, we propose a Transformer-Based Super-Resolution Network (TBSRN) containing a Self-Attention Module to extract sequential information, which is robust to tackle the texts in arbitrary orientations. In terms of character-level details, we propose a Position-Aware Module and a Content-Aware Module to highlight the position and the content of each character. By observing that some characters look indistinguishable in low-resolution conditions, we use a weighted cross-entropy loss to tackle this problem. We conduct extensive experiments, including text recognition with pre-trained recognizers and image quality evaluation, on TextZoom and several scene text recognition benchmarks to assess the super-resolution images. The experimental results show that our STT can indeed generate text-focused super-resolution images and outperform the existing methods in terms of recognition accuracy.

Scene text retrieval aims to localize and search all text instances from an image gallery, which are the same or similar with a given query text. Such a task is usually realized by matching a query text to the recognized words, outputted by an end-to-end scene text spotter. In this paper, we address this problem by directly learning a cross-modal similarity between a query text and each text instance from natural images. Specifically, we establish an end-to-end trainable network, jointly optimizing the procedures of scene text detection and cross-modal similarity learning. In this way, scene text retrieval can be simply performed by ranking the detected text instances with the learned similarity. Experiments on three benchmark datasets demonstrate our method consistently outperforms the state-of-the-art scene text spotting/retrieval approaches. In particular, the proposed framework of joint detection and similarity learning achieves significantly better performance than separated methods.

What scene elements, if any, are indispensable for recognizing a scene? We strive to answer this question through the lens of an end-to-end learning scheme. Our goal is to identify a collection of such pivotal elements, which we term as Scene Essence, to be those that would alter scene recognition if taken out from the scene. To this end, we devise a novel approach that learns to partition the scene objects into two groups, essential ones and minor ones, under the supervision that if only the essential ones are kept while the minor ones are erased in the input image, a scene recognizer would preserve its original prediction. Specifically, we introduce a learnable graph neural network (GNN) for labelling scene objects, based on which the minor ones are wiped off by an off-the-shelf image inpainter. The features of the inpainted image derived in this way, together with those learned from the GNN with the minor-object nodes pruned, are expected to fool the scene discriminator. Both subjective and objective evaluations on Places365, SUN397, and MIT67 datasets demonstrate that, the learned Scene Essence yields a visually plausible image that convincingly retains the original scene category.

We introduce the new task of dense captioning in RGB-D scans. As input, we assume a point cloud of a 3D scene; the expected output is the bounding boxes along with the descriptions for the underlying objects. To address the 3D object detecting and describing problem at the same time, we propose Scan2Cap, an end-to-end trained architecture, to detect objects in the input scene and generate the descriptions for all of them in natural language. We apply an attention-based captioning method to generate descriptive tokens while referring to the related components in the local context. To better handle the relative spatial relations between objects, a message passing graph module is applied to learn the relation features, which are later used in the captioning phase. On the recently proposed ScanRefer dataset, we show that our architecture can effectively localize and describe the 3D objects in the scene. It also outperforms the 2D-based methods on the 3D dense captioning task by a big margin.

Self-attention has the promise of improving computer vision systems due to parameter-independent scaling of receptive fields and content-dependent interactions, in contrast to parameter-dependent scaling and content-independent interactions of convolutions. Self-attention models have recently been shown to have encouraging improvements on accuracy-parameter trade-offs compared to baseline convolutional models such as ResNet-50. In this work, we develop self-attention models that can outperform not just the canonical baseline models, but even the high-performing convolutional models. We propose two extensions to self-attention that, in conjunction with a more efficient implementation of self-attention, improve the speed, memory usage, and accuracy of these models. We leverage these improvements to develop a new self-attention model family, HaloNets, which reach state-of-the-art accuracies on the parameter-limited setting of the ImageNet classification benchmark. In preliminary transfer learning experiments, we find that HaloNet models outperform much larger models and have better inference performance. On harder tasks such as object detection and instance segmentation, our simple local self-attention and convolutional hybrids show improvements over very strong baselines. These results mark another step in demonstrating the efficacy of self-attention models on settings traditionally dominated by convolutions.

We show that the YOLOv4 object detection neural network based on the CSP approach, scales both up and down and is applicable to small and large networks while maintaining optimal speed and accuracy. We propose a network scaling approach that modifies not only the depth, width, resolution, but also structure of the network. YOLOv4-large model achieves state-of-the-art results: 55.5% AP (73.4% AP50) for the MS COCO dataset at a speed of 16 FPS on Tesla V100, while with the test time augmentation, YOLOv4-large achieves 56.0% AP (73.3 AP50). To the best of our knowledge, this is currently the highest accuracy on the COCO dataset among any published work. The YOLOv4-tiny model achieves 22.0% AP (42.0% AP50) at a speed of 443 FPS on RTX 2080Ti, while by using TensorRT, batch size = 4 and FP16-precision the YOLOv4-tiny achieves 1774 FPS.

We consider the abstract relational reasoning task, which is commonly used as an intelligence test. Since some patterns have spatial rationales, while others are only semantic, we propose a multi-scale architecture that processes each query in multiple resolutions. We show that indeed different rules are solved by different resolutions and a combined multi-scale approach outperforms the existing state of the art in this task on all benchmarks by 5-54%. The success of our method is shown to arise from multiple novelties. First, it searches for relational patterns in multiple resolutions, which allows it to readily detect visual relations, such as location, in higher resolution, while allowing the lower resolution module to focus on semantic relations, such as shape type. Second, we optimize the reasoning network of each resolution proportionally to its performance, hereby we motivate each resolution to specialize on the rules for which it performs better than the others and ignore cases that are already solved by the other resolutions. Third, we propose a new way to pool information along the rows and the columns of the illustration-grid of the query. Our work also analyses the existing benchmarks, demonstrating that the RAVEN dataset selects the negative examples in a way that is easily exploited. We, therefore, propose a modified version of the RAVEN dataset, named RAVEN-FAIR. Our code and pretrained models are available at https://github.com/yanivbenny/MRNet.

Few-shot semantic segmentation (FSS) aims to segment unseen class objects given very few densely-annotated support images from the same class. Existing FSS methods find the query object by using support prototypes or by directly relying on heuristic multi-scale feature fusion. However, they fail to fully leverage the high-order appearance relationships between multi-scale features among the support-query image pairs, thus leading to an inaccurate localization of the query objects. To tackle the above challenge, we propose an end-to-end scale-aware graph neural network (SAGNN) by reasoning the cross-scale relations among the support-query images for FSS. Specifically, a scale-aware graph is first built by taking support-induced multi-scale query features as nodes and, meanwhile, each edge is modeled as the pairwise interaction of its connected nodes. By progressive message passing over this graph, SAGNN is capable of capturing cross-scale relations and overcoming object variations (e.g., appearance, scale and location), and can thus learn more precise node embeddings. This in turn enables it to predict more accurate foreground objects. Moreover, to make full use of the location relations across scales for the query image, a novel self-node collaboration mechanism is proposed to enrich the current node, which endows SAGNN the ability of perceiving different resolutions of the same objects. Extensive experiments on PASCAL-5i and COCO-20i show that SAGNN achieves state-of-the-art results.

We propose Scale-aware AutoAug to learn data augmentation policies for object detection. We define a new scale-aware search space, where both image- and box-level augmentations are designed for maintaining scale invariance. Upon this search space, we propose a new search metric, termed Pareto Scale Balance, to facilitate search with high efficiency. In experiments, Scale-aware AutoAug yields significant and consistent improvement on various object detectors (e.g., RetinaNet, Faster R-CNN, Mask R-CNN, and FCOS), even compared with strong multi-scale training baselines. Our searched augmentation policies are transferable to other datasets and box-level tasks beyond object detection (e.g., instance segmentation and keypoint estimation) to improve performance. The search cost is much less than previous automated augmentation approaches for object detection. It is notable that our searched policies have meaningful patterns, which intuitively provide valuable insight for human data augmentation design.

We propose a novel method for privacy-preserving training of deep neural networks leveraging public, out-domain data. While differential privacy (DP) has emerged as a mechanism to protect sensitive data in training datasets, its application to complex visual recognition tasks remains challenging. Traditional DP methods, such as Differentially-Private Stochastic Gradient Descent (DP-SGD), only perform well on simple datasets and shallow networks, while recent transfer learning-based DP methods often make unrealistic assumptions about the availability and distribution of public data. In this work, we argue that minimizing the number of trainable parameters is the key to improving the privacy-performance tradeoff of DP on complex visual recognition tasks. We also propose a novel transfer learning paradigm that finetunes a very sparse subnetwork with DP, inspired by this argument. We conduct extensive experiments and ablation studies on two visual recognition tasks: CIFAR-100 -> CIFAR-10 (standard DP setting) and the CD-FSL challenge (few-shot, multiple levels of domain shifts) and demonstrate competitive experimental performance.

Quantifying the importance of each training point to a learning task is a fundamental problem in machine learning and the estimated importance scores have been leveraged to guide a range of data workflows such as data summarization and domain adaption. One simple idea is to use the leave-one-out error of each training point to indicate its importance. Recent work has also proposed to use the Shapley value, as it defines a unique value distribution scheme that satisfies a set of appealing properties. However, calculating Shapley values is often expensive, which limits its applicability in real-world applications at scale. Multiple heuristics to improve the scalability of calculating Shapley values have been proposed recently, with the potential risk of compromising their utility in real-world applications. How well do existing data quantification methods perform on existing workflows? How do these methods compare with each other, empirically and theoretically? Must we sacrifice scalability for the utility in these workflows when using these methods? In this paper, we conduct a novel theoretical analysis comparing the utility of different importance quantification methods, and report extensive experimental studies on settings such as noisy label detection, watermark removal, data summarization, data acquisition, and domain adaptation on existing and proposed workflows. We show that Shapley value approximation based on a KNN surrogate over pre-trained feature embeddings obtains comparable utility with existing algorithms while achieving significant scalability improvement, often by orders of magnitude. Our theoretical analysis also justifies its advantage over the leave-one-out error. The code is available at https://github.com/AI-secure/Shapley-Study.

In this paper, we propose a novel task for saliency-guided image translation, with the goal of image-to-image translation conditioned on the user specified saliency map. To address this problem, we develop a novel Generative Adversarial Network (GAN)-based model, called SalG-GAN. Given the original image and target saliency map, SalG-GAN can generate a translated image that satisfies the target saliency map. In SalG-GAN, a disentangled representation framework is proposed to encourage the model to learn diverse translations for the same target saliency condition. A saliency-based attention module is introduced as a special attention mechanism for facilitating the developed structures of saliency-guided generator, saliency cue encoder and saliency-guided global and local discriminators. Furthermore, we build a synthetic dataset and a real-world dataset with labeled visual attention for training and evaluating our SalG-GAN. The experimental results over both datasets verify the effectiveness of our model for saliency-guided image translation.

Safe local motion planning for autonomous driving in dynamic environments requires forecasting how the scene evolves. Practical autonomy stacks adopt a semantic object-centric representation of a dynamic scene and build object detection, tracking, and prediction modules to solve forecasting. However, training these modules comes at an enormous human cost of manually annotated objects across frames. In this work, we explore future freespace as an alternative representation to support motion planning. Our key intuition is that it is important to avoid straying into occupied space regardless of what is occupying it. Importantly, computing ground-truth future freespace is annotation-free. First, we explore freespace forecasting as a self-supervised learning task. We then demonstrate how to use forecasted freespace to identify collision-prone plans from off-the-shelf motion planners. Finally, we propose future freespace as an additional source of annotation-free supervision. We demonstrate how to integrate such supervision into the learning-based planners. Experimental results on nuScenes and CARLA suggest both approaches lead to a significant reduction in collision rates.

Traffic event cognition and reasoning in videos is an important task that has a wide range of applications in intelligent transportation, assisted driving, and autonomous vehicles. In this paper, we create a novel dataset, SUTD-TrafficQA (Traffic Question Answering), which takes the form of video QA based on the collected 10,080 in-the-wild videos and annotated 62,535 QA pairs, for benchmarking the cognitive capability of causal inference and event understanding models in complex traffic scenarios. Specifically, we propose 6 challenging reasoning tasks corresponding to various traffic scenarios, so as to evaluate the reasoning capability over different kinds of complex yet practical traffic events. Moreover, we propose Eclipse, a novel Efficient glimpse network via dynamic inference, in order to achieve computation-efficient and reliable video reasoning. The experiments show that our method achieves superior performance while reducing the computation cost significantly.

We present STaR, a novel method that performs Self-supervised Tracking and Reconstruction of dynamic scenes with rigid motion from multi-view RGB videos without any manual annotation. Recent work has shown that neural networks are surprisingly effective at the task of compressing many views of a scene into a learned function which maps from a viewing ray to an observed radiance value via volume rendering. Unfortunately, these methods lose all their predictive power once any object in the scene has moved. In this work, we explicitly model rigid motion of objects in the context of neural representations of radiance fields. We show that without any additional human specified supervision, we can reconstruct a dynamic scene with a single rigid object in motion by simultaneously decomposing it into its two constituent parts and encoding each with its own neural representation. We achieve this by jointly optimizing the parameters of two neural radiance fields and a set of rigid poses which align the two fields at each frame. On both synthetic and real world datasets, we demonstrate that our method can render photorealistic novel views, where novelty is measured on both spatial and temporal axes. Our factored representation furthermore enables animation of unseen object motion.

Boosting performance of the offline trained siamese trackers is getting harder nowadays since the fixed information of the template cropped from the first frame has been almost thoroughly mined, but they are poorly capable of resisting target appearance changes. Existing trackers with template updating mechanisms rely on time-consuming numerical optimization and complex hand-designed strategies to achieve competitive performance, hindering them from real-time tracking and practical applications. In this paper, we propose a novel tracking framework built on top of a space-time memory network that is competent to make full use of historical information related to the target for better adapting to appearance variations during tracking. Specifically, a novel memory mechanism is introduced, which stores the historical information of the target to guide the tracker to focus on the most informative regions in the current frame. Furthermore, the pixel-level similarity computation of the memory network enables our tracker to generate much more accurate bounding boxes of the target. Extensive experiments and comparisons with many competitive trackers on challenging large-scale benchmarks, OTB-2015, TrackingNet, GOT-10k, LaSOT, UAV123, and VOT2018, show that, without bells and whistles, our tracker outperforms all previous state-of-the-art real-time methods while running at 37 FPS. The code is available at https://github.com/fzh0917/STMTrack.