ICCV 2021 Papers

Examples that are close to the decision boundary---that we term hard examples, are essential to shape accurate classifiers. Extracting confident examples has been widely studied in the community of learning with noisy labels. However, it remains elusive how to extract hard confident examples from the noisy training data. In this paper, we propose a deep learning paradigm to solve this problem, which is built on the memorization effect of deep neural networks that they would first learn simple patterns, i.e., which are defined by these shared by multiple training examples. To extract hard confident examples that contain non-simple patterns and are entangled with the inaccurately labeled examples, we borrow the idea of momentum from physics. Specifically, we alternately update the confident examples and refine the classifier. Note that the extracted confident examples in the previous round can be exploited to learn a better classifier and that the better classifier will help identify better (and hard) confident examples. We call the approach the "Momentum of Memorization" (Me-Momentum). Empirical results on benchmark-simulated and real-world label-noise data illustrate the effectiveness of Me-Momentum for extracting hard confident examples, leading to better classification performance.

This paper considers matching images of low-light scenes, aiming to widen the frontier of SfM and visual SLAM applications. Recent image sensors can record the brightness of scenes with more than eight-bit precision, available in their RAW-format image. We are interested in making full use of such high-precision information to match extremely low-light scene images that conventional methods cannot handle. For extreme low-light scenes, even if some of their brightness information exists in the RAW format images' low bits, the standard raw image processing fails to utilize them properly. As was recently shown by Chen et al., CNNs can learn to produce images with a natural appearance from such RAW-format images. To consider if and how well we can utilize such information stored in RAW-format images for image matching, we have created a new dataset named MID (matching in the dark). Using it, we experimentally evaluated combinations of eight image-enhancing methods and eleven image matching methods consisting of classical/neural local descriptors and classical/neural initial point-matching methods. The results show the advantage of using the RAW-format images and the strengths and weaknesses of the above component methods. They also imply there is room for further research.

We propose a manifold matching approach to generative models which includes a distribution generator (or data generator) and a metric generator. In our framework, we view the real data set as some manifold embedded in a high-dimensional Euclidean space. The distribution generator aims at generating samples that follow some distribution condensed around the real data manifold. It is achieved by matching two sets of points using their geometric shape descriptors, such as centroid and p-diameter, with learned distance metric; the metric generator utilizes both real data and generated samples to learn a distance metric which is close to some intrinsic geodesic distance on the real data manifold. The produced distance metric is further used for manifold matching. The two networks learn simultaneously during the training process. We apply the approach on both unsupervised and supervised learning tasks: in unconditional image generation task, the proposed method obtains competitive results compared with existing generative models; in super-resolution task, we incorporate the framework in perception-based models and improve visual qualities by producing samples with more natural textures. Experiments and analysis demonstrate the feasibility and effectiveness of the proposed framework.

Most existing style transfer methods follow the assumption that styles can be represented with global statistics (e.g., Gram matrices or covariance matrices), and thus address the problem by forcing the output and style images to have similar global statistics. An alternative is the assumption of local style patterns, where algorithms are designed to swap similar local features of content and style images. However, the limitation of these existing methods is that they neglect the semantic structure of the content image which may lead to corrupted content structure in the output. In this paper, we make a new assumption that image features from the same semantic region form a manifold and an image with multiple semantic regions follows a multi-manifold distribution. Based on this assumption, the style transfer problem is formulated as aligning two multi-manifold distributions and a Manifold Alignment based Style Transfer (MAST) framework is proposed. The proposed framework allows semantically similar regions between the output and the style image share similar style patterns. Moreover, the proposed manifold alignment method is flexible to allow user editing or using semantic segmentation maps as guidance for style transfer. To allow the method to be applicable to photorealistic style transfer, we propose a new adaptive weight skip connection network structure to preserve the content details. Extensive experiments verify the effectiveness of the proposed framework for both artistic and photorealistic style transfer. Code is available at https://github.com/NJUHuoJing/MAST.

We present an efficient approximate message passing solver for the lifted disjoint paths problem (LDP), a natural but NP-hard model for multiple object tracking (MOT). Our tracker scales to very large instances that come from long and crowded MOT sequences. Our approximate solver enables us to process the MOT15/16/17 benchmarks without sacrificing solution quality and allows for solving MOT20, which has been out of reach up to now for LDP solvers due to its size and complexity. On all these four standard MOT benchmarks we achieve performance comparable or better than current state-of-the-art methods including a tracker based on an optimal LDP solver.

Multi-view projection methods have demonstrated their ability to reach state-of-the-art performance on 3D shape recognition. Those methods learn different ways to aggregate information from multiple views. However, the camera view-points for those views tend to be heuristically set and fixed for all shapes. To circumvent the lack of dynamism of current multi-view methods, we propose to learn those view-points. In particular, we introduce the Multi-View Transformation Network (MVTN) that regresses optimal view-points for 3D shape recognition, building upon advances in differentiable rendering. As a result, MVTN can be trained end-to-end along with any multi-view network for 3D shape classification. We integrate MVTN in a novel adaptive multi-view pipeline that can render either 3D meshes or point clouds. MVTN exhibits clear performance gains in the tasks of 3D shape classification and 3D shape retrieval without the need for extra training supervision. In these tasks, MVTN achieves state-of-the-art performance on ModelNet40, ShapeNet Core55, and the most recent and realistic ScanObjectNN dataset (up to 6 % improvement). Interestingly, we also show that MVTN can provide network robustness against rotation and occlusion in the 3D domain.

We present MVSNeRF, a novel neural rendering approach that can efficiently reconstruct neural radiance fields for view synthesis. Unlike prior works on neural radiance fields that consider per-scene optimization on densely captured images, we propose a generic deep neural network that can reconstruct radiance fields from only three nearby input views via fast network inference. Our approach leverages plane-swept cost volumes (widely used in multi-view stereo) for geometry-aware scene reasoning, and combines this with physically based volume rendering for neural radiance field reconstruction. We train our network on real objects in the DTU dataset, and test it on three different datasets to evaluate its effectiveness and generalizability. Our approach can generalize across scenes (even indoor scenes, completely different from our training scenes of objects) and generate realistic view synthesis results using only three input images, significantly outperforming concurrent works on generalizable radiance field reconstruction. Moreover, if dense images are captured, our estimated radiance field representation can be easily fine-tuned; this leads to fast per-scene reconstruction with higher rendering quality and substantially less optimization time than NeRF.

Image quality assessment (IQA) is an important research topic for understanding and improving visual experience. The current state-of-the-art IQA methods are based on convolutional neural networks (CNNs). The performance of CNN-based models is often compromised by the fixed shape constraint in batch training. To accommodate this, the input images are usually resized and cropped to a fixed shape, causing image quality degradation. To address this, we design a multi-scale image quality Transformer (MUSIQ) to process native resolution images with varying sizes and aspect ratios. With a multi-scale image representation, our proposed method can capture image quality at different granularities. Furthermore, a novel hash-based 2D spatial embedding and a scale embedding is proposed to support the positional embedding in the multi-scale representation. Experimental results verify that our method can achieve state-of-the-art performance on multiple large scale IQA datasets such as PaQ-2-PiQ, SPAQ and KonIQ-10k.

Retrieving occlusion relation among objects in a single image is challenging due to sparsity of boundaries in image. We observe two key issues in existing works: firstly, lack of an architecture which can exploit the limited amount of coupling in the decoder stage between the two subtasks, namely occlusion boundary extraction and occlusion orientation prediction, and secondly, improper representation of occlusion orientation. In this paper, we propose a novel architecture called Occlusion-shared and Path-separated Network (OPNet), which solves the first issue by exploiting rich occlusion cues in shared high-level features and structured spatial information in task-specific low-level features. We then design a simple but effective orthogonal occlusion representation (OOR) to tackle the second issue. Our method surpasses the state-of-the-art methods by 6.1%/8.3% Boundary-AP and 6.5%/10% Orientation-AP on standard PIOD/BSDS ownership datasets. Code is available at https://github.com/fengpanhe/MT-ORL.

Human motion prediction is a challenging task due to the stochasticity and aperiodicity of future poses. Recently, graph convolutional network has been proven to be very effective to learn dynamic relations among pose joints, which is helpful for pose prediction. On the other hand, one can abstract a human pose recursively to obtain a set of poses at multiple scales. With the increase of the abstraction level, the motion of the pose becomes more stable, which benefits pose prediction too. In this paper, we propose a novel Multi-Scale Residual Graph Convolution Network (MSR-GCN) for human pose prediction task in the manner of end-to-end. The GCNs are used to extract features from fine to coarse scale and then from coarse to fine scale. The extracted features at each scale are then combined and decoded to obtain the residuals between the input and target poses. Intermediate supervisions are imposed on all the predicted poses, which enforces the network to learn more representative features. Our proposed approach is evaluated on two standard benchmark datasets, i.e., the Human3.6M dataset and the CMU Mocap dataset. Experimental results demonstrate that our method outperforms the state-of-the-art approaches. Code and pre-trained models are available at https://github.com/Droliven/MSRGCN.

Deep learning-based methods for video pedestrian detection and tracking require large volumes of training data to achieve good performance. However, data acquisition in crowded public environments raises data privacy concerns -- we are not allowed to simply record and store data without the explicit consent of all participants. Furthermore, the annotation of such data for computer vision applications usually requires a substantial amount of manual effort, especially in the video domain. Labeling instances of pedestrians in highly crowded scenarios can be challenging even for human annotators and may introduce errors in the training data. In this paper, we study how we can advance different aspects of multi-person tracking using solely synthetic data. To this end, we generate MOTSynth, a large, highly diverse synthetic dataset for object detection and tracking using a rendering game engine. Our experiments show that MOTSynth can be used as a replacement for real data on tasks such as pedestrian detection, re-identification, segmentation, and tracking.

Benefiting from the excellent performance of Siamese-based trackers, huge progress on 2D visual tracking has been achieved. However, 3D visual tracking is still under-explored. Inspired by the idea of Hough voting in 3D object detection, in this paper, we propose a Multi-level Voting Siamese Network (MLVSNet) for 3D visual tracking from outdoor point cloud sequences. To deal with sparsity in outdoor 3D point clouds, we propose to perform Hough voting on multi-level features to get more vote centers and retain more useful information, instead of voting only on the final level feature as in previous methods. We also design an efficient and lightweight Target-Guided Attention (TGA) module to transfer the target information and highlight the target points in the search area. Moreover, we propose a Vote-cluster Feature Enhancement (VFE) module to exploit the relationships between different vote clusters. Extensive experiments on the 3D tracking benchmark of KITTI dataset demonstrate that our MLVSNet outperforms state-of-the-art methods with significant margins. Code will be available at https://github.com/CodeWZT/MLVSNet.

In this paper, we propose MINE to perform novel view synthesis and depth estimation via dense 3D reconstruction from a single image. Our approach is a continuous depth generalization of the Multiplane Images (MPI) by introducing the NEural radiance fields (NeRF). Given a single image as input, MINE predicts a 4-channel image (RGB and volume density) at arbitrary depth values to jointly reconstruct the camera frustum and fill in occluded contents. The reconstructed and inpainted frustum can then be easily rendered into novel RGB or depth views using differentiable rendering. Extensive experiments on RealEstate10K, KITTI and Flowers Light Fields show that our MINE outperforms state-of-the-art by a large margin in novel view synthesis. We also achieve competitive results in depth estimation on iBims-1 and NYU-v2 without annotated depth supervision. Our source code is available at https://github.com/vincentfung13/MINE

Frame sampling is a fundamental problem in video action recognition due to the essential redundancy in time and limited computation resources. The existing sampling strategy often employs a fixed frame selection and lacks the flexibility to deal with complex variations in videos. In this paper, we present a simple, sparse, and explainable frame sampler, termed as Motion-Guided Sampler (MGSampler). Our basic motivation is that motion is an important and universal signal that can drive us to adaptively select frames from videos. Accordingly, we propose two important properties in our MGSampler design: motion sensitive and motion uniform. First, we present two different motion representations to enable us to efficiently distinguish the motion-salient frames from the background. Then, we devise a motion-uniform sampling strategy based on the cumulative motion distribution to ensure the sampled frames evenly cover all the important segments with high motion salience. Our MGSampler yields a new principled and holistic sample scheme, that could be incorporated into any existing video architecture. Experiments on five benchmarks demonstrate the effectiveness of our MGSampler over previous fixed sampling strategies, and its generalization power across different backbones, video models, and datasets.

We introduce MGNet, a multi-task framework for monocular geometric scene understanding. We define monocular geometric scene understanding as the combination of two known tasks: Panoptic segmentation and self-supervised monocular depth estimation. Panoptic segmentation captures the full scene not only semantically, but also on an instance basis. Self-supervised monocular depth estimation uses geometric constraints derived from the camera measurement model in order to measure depth from monocular video sequences only. To the best of our knowledge, we are the first to propose the combination of these two tasks in one single model. Our model is designed with focus on low latency to provide fast inference in real-time on a single consumer-grade GPU. During deployment, our model produces dense 3D point clouds with instance aware semantic labels from single high-resolution camera images. We evaluate our model on two popular autonomous driving benchmarks, i.e., Cityscapes and KITTI, and show competitive performance among other real-time capable methods. Source code is available at https://github.com/markusschoen/MGNet.

Pedestrian trajectory prediction is challenging due to its uncertain and multimodal nature. While generative adversarial networks can learn a distribution over future trajectories, they tend to predict out-of-distribution samples when the distribution of future trajectories is a mixture of multiple, possibly disconnected modes. To address this issue, we propose a multi-generator model for pedestrian trajectory prediction. Each generator specializes in learning a distribution over trajectories routing towards one of the primary modes in the scene, while a second network learns a categorical distribution over these generators, conditioned on the dynamics and scene input. This architecture allows us to effectively sample from specialized generators and to significantly reduce the out-of-distribution samples compared to single generator methods.

Weakly supervised salient object detection (WSOD) targets to train a CNNs-based saliency network using only low-cost annotations. Existing WSOD methods take various techniques to pursue single "high-quality" pseudo label from low-cost annotations and then develop their saliency networks. Though these methods have achieved good performance, the generated single label is inevitably affected by adopted refinement algorithms and shows prejudiced characteristics which further influence the saliency networks. In this work, we introduce a new multiple-pseudo label framework to integrate more comprehensive and accurate saliency cues from multiple labels, avoiding the aforementioned problem. Specifically, we propose a multi-filer directive network (MFNet) including a saliency network as well as multiple directive filters. The directive filter (DF) is designed to extract and filter more accurate saliency cues from the noisy pseudo labels. The multiple accurate cues from multiple DFs are then simultaneously propagated to the saliency network with a multi-guidance loss. Extensive experiments on five datasets over four metrics demonstrate that our method outperforms all the existing congeneric methods. Moreover, it is also worth noting that our framework is flexible enough to apply to existing methods and improve their performance.

Inspired by human visual attention, we propose a novel inverse reinforcement learning formulation using Maximum Entropy Deep Inverse Reinforcement Learning (MEDIRL) for predicting the visual attention of drivers in accident-prone situations. MEDIRL predicts fixation locations that lead to maximal rewards by learning a task-sensitive reward function from eye fixation patterns recorded from attentive drivers. Additionally, we introduce EyeCar, a new driver attention dataset in accident-prone situations. We conduct comprehensive experiments to evaluate our proposed model on three common benchmarks: (DR(eye)VE, BDD-A, DADA-2000), and our EyeCar dataset. Results indicate that MEDIRL outperforms existing models for predicting attention and achieves state-of-the-art performance. We present extensive ablation studies to provide more insights into different features of our proposed model.

Point completion refers to completing the missing geometries of an object from incomplete observations. Main-stream methods predict the missing shapes by decoding a global feature learned from the input point cloud, which often leads to deficient results in preserving topology consistency and surface details. In this work, we present ME-PCN, a point completion network that leverages `emptiness' in 3D shape space. Given a single depth scan, previous methods often encode the occupied partial shapes while ignoring the empty regions (e.g. holes) in depth maps. In contrast, we argue that these `emptiness' clues indicate shape boundaries that can be used to improve topology representation and detail granularity on surfaces. Specifically, our ME-PCN encodes both the occupied point cloud and the neighboring `empty points'. It estimates coarse-grained but complete and reasonable surface points in the first stage, followed by a refinement stage to produce fine-grained surface details. Comprehensive experiments verify that our ME-PCN presents better qualitative and quantitative performance against the state-of-the-art. Besides, we further prove that our `emptiness' design is lightweight and easy to embed in existing methods, which shows consistent effectiveness in improving the CD and EMD scores.