CVPR 2021 论文列表

Localization of anatomical landmarks to perform two-dimensional measurements in echocardiography is part of routine clinical workflow in cardiac disease diagnosis. Automatic localization of those landmarks is highly desirable to improve workflow and reduce interobserver variability. Training a machine learning framework to perform such localization is hindered given the sparse nature of gold standard labels; only few percent of cardiac cine series frames are normally manually labeled for clinical use. In this paper, we propose a new end-to-end reciprocal detection and tracking model that is specifically designed to handle the sparse nature of echocardiography labels. The model is trained using few annotated frames across the entire cardiac cine sequence to generate consistent detection and tracking of landmarks, and an adversarial training for the model is proposed to take advantage of these annotated frames. The superiority of the proposed reciprocal model is demonstrated using a series of experiments.

A set of cameras with fisheye lenses have been used to capture a wide field of view. The traditional scan-line stereo algorithms based on epipolar geometry are directly inapplicable to this non-pinhole camera setup due to optical characteristics of fisheye lenses; hence, existing complete 360-deg. RGB-D imaging systems have rarely achieved real-time performance yet. In this paper, we introduce an efficient sphere-sweeping stereo that can run directly on multiview fisheye images without requiring additional spherical rectification. Our main contributions are: First, we introduce an adaptive spherical matching method that accounts for each input fisheye camera's resolving power concerning spherical distortion. Second, we propose a fast inter-scale bilateral cost volume filtering method that refines distance in noisy and textureless regions with the optimal complexity of O(n). It enables real-time dense distance estimation while preserving edges. Lastly, the fisheye color and distance images are seamlessly combined into a complete 360-deg. RGB-D image via fast inpainting of the dense distance map. We demonstrate an embedded 360-deg. RGB-D imaging prototype composed of a mobile GPU and four fisheye cameras. Our prototype is capable of capturing complete 360-deg. RGB-D videos with a resolution of two megapixels at 29 fps. Results demonstrate that our real-time method outperforms traditional omnidirectional stereo and learning-based omnidirectional stereo in terms of accuracy and performance.

We propose a novel real-time selfie video stabilization method. Our method is completely automatic and runs at 26 fps. We use a 1D linear convolutional network to directly infer the rigid moving least squares warping which implicitly balances between the global rigidity and local flexibility. Our network structure is specifically designed to stabilize the background and foreground at the same time, while providing optional control of stabilization focus (relative importance of foreground vs. background) to the users. To train our network, we collect a selfie video dataset with 1005 videos, which is significantly larger than previous selfie video datasets. We also propose a grid approximation to the rigid moving least squares that enables the real-time frame warping. Our method is fully automatic and produces visually and quantitatively better results than previous real-time general video stabilization methods. Compared to previous offline selfie video methods, our approach produces comparable quality with a speed improvement of orders of magnitude. Our code and selfie video dataset is available at https://github.com/jiy173/selfievideostabilization.

We introduce a real-time, high-resolution background replacement technique which operates at 30fps in 4K resolution, and 60fps for HD on a modern GPU. Our technique is based on background matting, where an additional frame of the background is captured and used to inform the alpha matte and the foreground layer. The main challenge is to compute a high-quality alpha matte, preserving strand-level hair details, while processing high-resolution images in real-time. To achieve this goal, we employ two neural networks; the base network computes a low-resolution result which is refined by a second network operating at high-resolution on selective patches. We introduce two large-scale video and image matting datasets: VideoMatte240K and PhotoMatte13K/85. Our approach yields higher quality results compared to the previous state-of-the-art in background matting, while simultaneously yielding a dramatic boost in both speed and resolution.

The objective of this work is to annotate sign instances across a broad vocabulary in continuous sign language. We train a Transformer model to ingest a continuous signing stream and output a sequence of written tokens on a large-scale collection of signing footage with weakly-aligned subtitles. We show that through this training it acquires the ability to attend to a large vocabulary of sign instances in the input sequence, enabling their localisation. Our contributions are as follows: (1) we demonstrate the ability to leverage large quantities of continuous signing videos with weakly-aligned subtitles to localise signs in continuous sign language; (2) we employ the learned attention to automatically generate hundreds of thousands of annotations for a large sign vocabulary; (3) we collect a set of 37K manually verified sign instances across a vocabulary of 950 sign classes to support our study of sign language recognition; (4) by training on the newly annotated data from our method, we outperform the prior state of the art on the BSL-1K sign language recognition benchmark.

Linguistic knowledge is of great benefit to scene text recognition. However, how to effectively model linguistic rules in end-to-end deep networks remains a research challenge. In this paper, we argue that the limited capacity of language models comes from: 1) implicitly language modeling; 2) unidirectional feature representation; and 3) language model with noise input. Correspondingly, we propose an autonomous, bidirectional and iterative ABINet for scene text recognition. Firstly, the autonomous suggests to block gradient flow between vision and language models to enforce explicitly language modeling. Secondly, a novel bidirectional cloze network (BCN) as the language model is proposed based on bidirectional feature representation. Thirdly, we propose an execution manner of iterative correction for language model which can effectively alleviate the impact of noise input. Additionally, based on the ensemble of iterative predictions, we propose a self-training method which can learn from unlabeled images effectively. Extensive experiments indicate that ABINet has superiority on low-quality images and achieves state-of-the-art results on several mainstream benchmarks. Besides, the ABINet trained with ensemble self-training shows promising improvement in realizing human-level recognition.

An effective and efficient architecture performance evaluation scheme is essential for the success of Neural Architecture Search (NAS). To save computational cost, most of existing NAS algorithms often train and evaluate intermediate neural architectures on a small proxy dataset with limited training epochs. But it is difficult to expect an accurate performance estimation of an architecture in such a coarse evaluation way. This paper advocates a new neural architecture evaluation scheme, which aims to determine which architecture would perform better instead of accurately predict the absolute architecture performance. Therefore, we propose a relativistic architecture performance predictor in NAS (ReNAS). We encode neural architectures into feature tensors, and further refining the representations with the predictor. The proposed relativistic performance predictor can be deployed in discrete searching methods to search for the desired architectures without additional evaluation. Experimental results on NAS-Bench-101 dataset suggests that, sampling 424 (0.1% of the entire search space) neural architectures and their corresponding validation performance is already enough for learning an accurate architecture performance predictor. The accuracies of our searched neural architectures on NAS-Bench-101 and NAS-Bench-201 datasets are higher than that of the state-of-the-art methods and show the priority of the proposed method.

Image-to-image (I2I) translation methods based on generative adversarial networks (GANs) typically suffer from overfitting when limited training data is available. In this work, we propose a data augmentation method (ReMix) to tackle this issue. We interpolate training samples at the feature level and propose a novel content loss based on the perceptual relations among samples. The generator learns to translate the in-between samples rather than memorizing the training set, and thereby forces the discriminator to generalize. The proposed approach effectively reduces the ambiguity of generation and renders content-preserving results. The ReMix method can be easily incorporated into existing GAN models with minor modifications. Experimental results on numerous tasks demonstrate that GAN models equipped with the ReMix method achieve significant improvements.

Recently, object detection in aerial images has gained much attention in computer vision. Different from objects in natural images, aerial objects are often distributed with arbitrary orientation. Therefore, the detector requires more parameters to encode the orientation information, which are often highly redundant and inefficient. Moreover, as ordinary CNNs do not explicitly model the orientation variation, large amounts of rotation augmented data is needed to train an accurate object detector. In this paper, we propose a Rotation-equivariant Detector (ReDet) to address these issues, which explicitly encodes rotation equivariance and rotation invariance. More precisely, we incorporate rotation-equivariant networks into the detector to extract rotation-equivariant features, which can accurately predict the orientation and lead to a huge reduction of model size. Based on the rotation-equivariant features, we also present Rotation-invariant RoI Align (RiRoI Align), which adaptively extracts rotation-invariant features from equivariant features according to the orientation of RoI. Extensive experiments on several challenging aerial image datasets DOTA-v1.0, DOTA-v1.5 and HRSC2016, show that our method can achieve state-of-the-art performance on the task of aerial object detection. Compared with previous best results, our ReDet gains 1.2, 3.5 and 2.6 mAP on DOTA-v1.0, DOTA-v1.5 and HRSC2016 respectively while reducing the number of parameters by 60% (313 Mb vs. 121 Mb). The code is available at: https://github.com/csuhan/ReDet.

Point cloud registration is a common step in many 3D computer vision tasks such as object pose estimation, where a 3D model is aligned to an observation. Classical registration methods generalize well to novel domains but fail when given a noisy observation or a bad initialization. Learning-based methods, in contrast, are more robust but lack in generalization capacity. We propose to consider iterative point cloud registration as a reinforcement learning task and, to this end, present a novel registration agent (ReAgent). We employ imitation learning to initialize its discrete registration policy based on a steady expert policy. Integration with policy optimization, based on our proposed alignment reward, further improves the agent's registration performance. We compare our approach to classical and learning-based registration methods on both ModelNet40 (synthetic) and ScanObjectNN (real data) and show that our ReAgent achieves state-of-the-art accuracy. The lightweight architecture of the agent, moreover, enables reduced inference time as compared to related approaches.

ImageNet has been the most popular image classification benchmark, but it is also the one with a significant level of label noise. Recent studies have shown that many samples contain multiple classes, despite being assumed to be a single-label benchmark. They have thus proposed to turn ImageNet evaluation into a multi-label task, with exhaustive multi-label annotations per image. However, they have not fixed the training set, presumably because of a formidable annotation cost. We argue that the mismatch between single-label annotations and effectively multi-label images is equally, if not more, problematic in the training setup, where random crops are applied. With the single-label annotations, a random crop of an image may contain an entirely different object from the ground truth, introducing noisy or even incorrect supervision during training. We thus re-label the ImageNet training set with multi-labels. We address the annotation cost barrier by letting a strong image classifier, trained on an extra source of data, generate the multi-labels. We utilize the pixel-wise multi-label predictions before the final pooling layer, in order to exploit the additional location-specific supervision signals. Training on the re-labeled samples results in improved model performances across the board. ResNet-50 attains the top-1 accuracy of 78.9% on ImageNet with our localized multi-labels, which can be further boosted to 80.2% with the CutMix regularization. We show that the models trained with localized multi-labels also outperform the baselines on transfer learning to object detection and instance segmentation tasks, and various robustness benchmarks. The re-labeled ImageNet training set, pre-trained weights, and the source code are available at https://github.com/naver-ai/relabel_imagenet.

This paper is concerned with ranking many pre-trained deep neural networks (DNNs), called checkpoints, for the transfer learning to a downstream task. Thanks to the broad use of DNNs, we may easily collect hundreds of checkpoints from various sources. Which of them transfers the best to our downstream task of interest? Striving to answer this question thoroughly, we establish a neural checkpoint ranking benchmark (NeuCRaB) and study some intuitive ranking measures. These measures are generic, applying to the checkpoints of different output types without knowing how the checkpoints are pre-trained on which dataset. They also incur low computation cost, making them practically meaningful. Our results suggest that the linear separability of the features extracted by the checkpoints is a strong indicator of transferability. We also arrive at a new ranking measure, NLEEP, which gives rise to the best performance in the experiments.

Modern object detection approaches cast detecting objects as optimizing two subtasks of classification and localization simultaneously. Existing methods often learn the classification task by optimizing each proposal separately and neglect the relationship among different proposals. Such detection paradigm also encounters the mismatch between classification and localization due to the inherent discrepancy of their optimization targets. In this work, we propose a ranking-based optimization algorithm for harmoniously learning to rank and localize proposals in lieu of the classification task. To this end, we comprehensively investigate three types of ranking constraints, i.e., global ranking, class-specific ranking and IoU-guided ranking losses. The global ranking loss encourages foreground samples to rank higher than background. The class-specific ranking loss ensures that positive samples rank higher than negative ones for each specific class. The IoU-guided ranking loss aims to align each pair of confidence scores with the associated pair of IoU overlap between two positive samples of a specific class. Our ranking constraints can sufficiently explore the relationships between samples from three different perspectives. They are easy-to-implement, compatible with mainstream detection frameworks and computation-free for inference. Experiments demonstrate that our RankDetNet consistently surpasses prior anchor-based and anchor-free baselines, e.g., improving RetinaNet baseline by 2.5% AP on the COCO test-dev set without bells and whistles. We also apply the proposed ranking constraints for 3D object detection and achieve improved performance, which further validates the superiority and generality of our method.

Scene recovery is a fundamental imaging task for several practical applications, e.g., video surveillance and autonomous vehicles, etc. To improve visual quality under different weather/imaging conditions, we propose a real-time light correction method to recover the degraded scenes in the cases of sandstorms, underwater, and haze. The heart of our work is that we propose an intensity projection strategy to estimate the transmission. This strategy is motivated by a straightforward rank-one transmission prior. The complexity of transmission estimation is O(N) where N is the size of the single image. Then we can recover the scene in real-time. Comprehensive experiments on different types of weather/imaging conditions illustrate that our method outperforms competitively several state-of-the-art imaging methods in terms of efficiency and robustness.

Real-time and high-performance 3D object detection is an attractive research direction in autonomous driving. Recent studies prefer point based or voxel based convolution for achieving high performance. However, these methods suffer from the unsatisfied efficiency or complex customized convolution, making them unsuitable for applications with real-time requirements. In this paper, we present an efficient and effective 3D object detection framework, named RangeIoUDet that uses the range image as input. Benefiting from the dense representation of the range image, RangeIoUDet is entirely constructed based on 2D convolution, making it possible to have a fast inference speed. This model learns pointwise features from the range image, which is then passed to a region proposal network for predicting 3D bounding boxes. We optimize the pointwise feature and the 3D box via the point-based IoU and box-based IoU supervision, respectively. The point-based IoU supervision is proposed to make the network better learn the implicit 3D information encoded in the range image. The 3D Hybrid GIoU loss is introduced to generate high-quality boxes while providing an accurate quality evaluation. Through the point-based IoU and the box-based IoU, RangeIoUDet outperforms all single-stage models on the KITTI dataset, while running at 45 FPS for inference. Experiments on the self-built dataset further prove its effectiveness on different LIDAR sensors and object categories.

Continual learning is a realistic learning scenario for AI models. Prevalent scenario of continual learning, however, assumes disjoint sets of classes as tasks and is less realistic rather artificial. Instead, we focus on 'blurry' task boundary; where tasks shares classes and is more realistic and practical. To address such task, we argue the importance of diversity of samples in an episodic memory. To enhance the sample diversity in the memory, we propose a novel memory management strategy based on per-sample classification uncertainty and data augmentation, named Rainbow Memory (RM). With extensive empirical validations on MNIST, CIFAR10, CIFAR100, and ImageNet datasets, we show that the proposed method significantly improves the accuracy in blurry continual learning setups, outperforming state of the arts by large margins despite its simplicity. Code and data splits will be available in https://github.com/clovaai/rainbow-memory.

Existing studies in weakly-supervised semantic segmentation (WSSS) using image-level weak supervision have several limitations: sparse object coverage, inaccurate object boundaries, and co-occurring pixels from non-target objects. To overcome these challenges, we propose a novel framework, namely Explicit Pseudo-pixel Supervision (EPS), which learns from pixel-level feedback by combining two weak supervisions; the image-level label provides the object identity via the localization map and the saliency map from the off-the-shelf saliency detection model offers rich boundaries. We devise a joint training strategy to fully utilize the complementary relationship between both information. Our method can obtain accurate object boundaries and discard co-occurring pixels, thereby significantly improving the quality of pseudo-masks. Experimental results show that the proposed method remarkably outperforms existing methods by resolving key challenges of WSSS and achieves the new state-of-the-art performance on both PASCAL VOC 2012 and MS COCO 2014 datasets. The code is available at https://github.com/halbielee/EPS.

While radar and video data can be readily fused at the detection level, fusing them at the pixel level is potentially more beneficial. This is also more challenging in part due to the sparsity of radar, but also because automotive radar beams are much wider than a typical pixel combined with a large baseline between camera and radar, which results in poor association between radar pixels and color pixel. A consequence is that depth completion methods designed for LiDAR and video fare poorly for radar and video. Here we propose a radar-to-pixel association stage which learns a mapping from radar returns to pixels. This mapping also serves to densify radar returns. Using this as a first stage, followed by a more traditional depth completion method, we are able to achieve image-guided depth completion with radar and video. We demonstrate performance superior to camera and radar alone on the nuScenes dataset. Our source code is available at https://github.com/longyunf/rc-pda.

Deploying deep convolutional neural networks on ultra-low power systems is challenging due to the extremely limited resources. Especially, the memory becomes a bottleneck as the systems put a hard limit on the size of on-chip memory. Because peak memory explosion in the lower layers is critical even in tiny models, the size of an input image should be reduced with sacrifice in accuracy. To overcome this drawback, we propose a novel Raster-Scanning Network, named RaScaNet, inspired by raster-scanning in image sensors. RaScaNet reads only a few rows of pixels at a time using a convolutional neural network and then sequentially learns the representation of the whole image using a recurrent neural network. The proposed method operates on an ultra-low power system without input size reduction; it requires 15.9-24.3x smaller peak memory and 5.3-12.9x smaller weight memory than the state-of-the-art tiny models. Moreover, RaScaNet fully exploits on-chip SRAM and cache memory of the system as the sum of the peak memory and the weight memory does not exceed 60 KB, improving the power efficiency of the system. In our experiments, we demonstrate the binary classification performance of RaScaNet on Visual Wake Words and Pascal VOC datasets.