CVPR 2021 Papers

Graph neural networks (GNNs) emerged recently as a standard toolkit for learning from data on graphs. Current GNN designing works depend on immense human expertise to explore different message-passing mechanisms, and require manual enumeration to determine the proper message-passing depth. Inspired by the strong searching capability of neural architecture search (NAS) in CNN, this paper proposes Graph Neural Architecture Search (GNAS) with novel-designed search space. The GNAS can automatically learn better architecture with the optimal depth of message passing on the graph. Specifically, we design Graph Neural Architecture Paradigm (GAP) with tree-topology computation procedure and two types of fine-grained atomic operations (feature filtering & neighbor aggregation) from message-passing mechanism to construct powerful graph network search space. Feature filtering performs adaptive feature selection, and neighbor aggregation captures structural information and calculates neighbors' statistics. Experiments show that our GNAS can search for better GNNs with multiple message-passing mechanisms and optimal message-passing depth. The searched network achieves remarkable improvement over state-of-the-art manual designed and search-based GNNs on five large-scale datasets at three classical graph tasks.

The majority of existing few-shot learning methods describe image relations with binary labels. However, such binary relations are insufficient to teach the network complicated real-world relations, due to the lack of decision smoothness. Furthermore, current few-shot learning models capture only the similarity via relation labels, but they are not exposed to class concepts associated with objects, which is likely detrimental to the classification performance due to underutilization of the available class labels. For instance, children learn the concept of tiger from a few of actual examples as well as from comparisons of tiger to other animals. Thus, we hypothesize that both similarity and class concept learning must be occurring simultaneously. With these observations at hand, we study the fundamental problem of simplistic class modeling in current few-shot learning methods. We rethink the relations between class concepts, and propose a novel Absolute-relative Learning paradigm to fully take advantage of label information to refine the image an relation representations in both supervised and unsupervised scenarios. Our proposed paradigm improves the performance of several state-of-the-art models on publicly available datasets.

Designing an efficient model within the limited computational cost is challenging. We argue the accuracy of a lightweight model has been further limited by the design convention: a stage-wise configuration of the channel dimensions, which looks like a piecewise linear function of the network stage. In this paper, we study an effective channel dimension configuration towards better performance than the convention. To this end, we empirically study how to design a single layer properly by analyzing the rank of the output feature. We then investigate the channel configuration of a model by searching network architectures concerning the channel configuration under the computational cost restriction. Based on the investigation, we propose a simple yet effective channel configuration that can be parameterized by the layer index. As a result, our proposed model following the channel parameterization achieves remarkable performance on ImageNet classification and transfer learning tasks including COCO object detection, COCO instance segmentation, and fine-grained classifications. Code and ImageNet pretrained models are available at https://github.com/clovaai/rexnet.

BiSeNet has been proved to be a popular two-stream network for real-time segmentation. However, its principle of adding an extra path to encode spatial information is time-consuming, and the backbones borrowed from pretrained tasks, e.g., image classification, may be inefficient for image segmentation due to the deficiency of task-specific design. To handle these problems, we propose a novel and efficient structure named Short-Term Dense Concatenate network (STDC network) by removing structure redundancy. Specifically, we gradually reduce the dimension of feature maps and use the aggregation of them for image representation, which forms the basic module of STDC network. In the decoder, we propose a Detail Aggregation module by integrating the learning of spatial information into low-level layers in single-stream manner. Finally, the low-level features and deep features are fused to predict the final segmentation results. Extensive experiments on Cityscapes and CamVid dataset demonstrate the effectiveness of our method by achieving promising trade-off between segmentation accuracy and inference speed. On Cityscapes, we achieve 71.9% mIoU on the test set with a speed of 250.4 FPS on NVIDIA GTX 1080Ti, which is 45.2% faster than the latest methods, and achieve 76.8% mIoU with 97.0 FPS while inferring on higher resolution images.

A practical low-light enhancement solution must be computationally fast, memory-efficient, and achieve a visually appealing restoration. Most of the existing methods target restoration quality and thus compromise on speed and memory requirements, raising concerns about their real-world deployability. We propose a new deep learning architecture for extreme low-light single image restoration, which is exceptionally lightweight, remarkably fast, and produces a restoration that is perceptually at par with state-of-the-art computationally intense models. To achieve this, we do most of the processing in the higher scale-spaces, skipping the intermediate-scales wherever possible. Also unique to our model is the potential to process all the scale-spaces concurrently, offering an additional 30% speedup without compromising the restoration quality. Pre-amplification of the dark raw-image is an important step in extreme low-light image enhancement. Most of the existing state-of-the-art methods need GT exposure value to estimate the pre-amplification factor, which is not practically feasible. Thus, we propose an amplifier module that estimates the amplification factor using only the input raw image and can be used "off-the-shelf"" with pre-trained models without any fine-tuning. We show that our model can restore an ultra-high-definition 4K resolution image in just 1sec on a CPU and at 32fps on a GPU and yet maintain a competitive restoration quality. We also show that our proposed model, without any fine-tuning, generalizes well to cameras not seen during training and to subsequent tasks such as object detection.

In this study, we propose a self-supervised video denoising method called ""restore-from-restored."" This method fine-tunes a pre-trained network by using a pseudo clean video during the test phase. The pseudo clean video is obtained by applying a noisy video to the baseline network. By adopting a fully convolutional neural network (FCN) as the baseline, we can improve video denoising performance without accurate optical flow estimation and registration steps, in contrast to many conventional video restoration methods, due to the translation equivariant property of the FCN. Specifically, the proposed method can take advantage of plentiful similar patches existing across multiple consecutive frames (i.e., patch-recurrence); these patches can boost the performance of the baseline network by a large margin. We analyze the restoration performance of the fine-tuned video denoising networks with the proposed self-supervision-based learning algorithm, and demonstrate that the FCN can utilize recurring patches without requiring accurate registration among adjacent frames. In our experiments, we apply the proposed method to state-of-the-art denoisers and show that our fine-tuned networks achieve a considerable improvement in denoising performance.

Recognition and reconstruction of residential floor plan drawings are important and challenging in design, decoration, and architectural remodeling fields. An automatic framework is provided that accurately recognizes the structure, type, and size of the room, and outputs vectorized 3D reconstruction results. Deep segmentation and detection neural networks are utilized to extract room structural information. Key points detection network and cluster analysis are utilized to calculate scales of rooms. The vectorization of room information is processed through an iterative optimization-based method. The system significantly increases accuracy and generalization ability, compared with existing methods. It outperforms other systems in floor plan segmentation and vectorization process, especially inclined wall detection.

While GANs have shown success in realistic image generation, the idea of using GANs for other tasks unrelated to synthesis is underexplored. Do GANs learn meaningful structural parts of objects during their attempt to reproduce those objects? In this work, we test this hypothesis and propose a simple and effective approach based on GANs for semantic part segmentation that requires as few as one label example along with an unlabeled dataset. Our key idea is to leverage a trained GAN to extract a pixel-wise representation from the input image and use it as feature vectors for a segmentation network. Our experiments demonstrate that this GAN-derived representation is "readily discriminative" and produces surprisingly good results that are comparable to those from supervised baselines trained with significantly more labels. We believe this novel repurposing of GANs underlies a new class of unsupervised representation learning, which can generalize to many other tasks.

Human actions are typically of combinatorial structures or patterns, i.e., subjects, objects, plus spatio-temporal interactions in between. Discovering such structures is therefore a rewarding way to reason about the dynamics of interactions and recognize the actions. In this paper, we introduce a new design of sub-graphs to represent and encode the discriminative patterns of each action in the videos. Specifically, we present MUlti-scale Sub-graph LEarning (MUSLE) framework that novelly builds space-time graphs and clusters the graphs into compact sub-graphs on each scale with respect to the number of nodes. Technically, MUSLE produces 3D bounding boxes, i.e., tubelets, in each video clip, as graph nodes and takes dense connectivity as graph edges between tubelets. For each action category, we execute online clustering to decompose the graph into sub-graphs on each scale through learning Gaussian Mixture Layer and select the discriminative sub-graphs as action prototypes for recognition. Extensive experiments are conducted on both Something-Something V1 & V2 and Kinetics-400 datasets, and superior results are reported when comparing to state-of-the-art methods. More remarkably, our MUSLE achieves to-date the best reported accuracy of 65.0% on Something-Something V2 validation set.

Although vanilla Convolutional Neural Network (CNN) based detectors can achieve satisfactory performance on fake face detection, we observe that the detectors tend to seek forgeries on a limited region of face, which reveals that the detectors is short of understanding of forgery. Therefore, we propose an attention-based data augmentation framework to guide detector refine and enlarge its attention. Specifically, our method tracks and occludes the Top-N sensitive facial regions, encouraging the detector to mine deeper into the regions ignored before for more representative forgery. Especially, our method is simple-to-use and can be easily integrated with various CNN models. Extensive experiments show that the detector trained with our method is capable to separately point out the representative forgery of fake faces generated by different manipulation techniques, and our method enables a vanilla CNN-based detector to achieve state-of-the-art performance without structure modification.

Batch Normalization (BatchNorm) has become the default component in modern neural networks to stabilize training. In BatchNorm, centering and scaling operations, along with mean and variance statistics, are utilized for feature standardization over the batch dimension. The batch dependency of BatchNorm enables stable training and better representation of the network, while inevitably ignores the representation differences among instances. We propose to add a simple yet effective feature calibration scheme into the centering and scaling operations of BatchNorm, enhancing the instance-specific representations with the negligible computational cost. The centering calibration strengthens informative features and reduces noisy features. The scaling calibration restricts the feature intensity to form a more stable feature distribution. Our proposed variant of BatchNorm, namely Representative BatchNorm, can be plugged into existing methods to boost the performance of various tasks such as classification, detection, and segmentation. The source code is available in http://mmcheng.net/rbn.

We introduce a weakly supervised method for representation learning based on aligning temporal sequences (e.g., videos) of the same process (e.g., human action). The main idea is to use the global temporal ordering of latent correspondences across sequence pairs as a supervisory signal. In particular, we propose a loss based on scoring the optimal sequence alignment to train an embedding network. Our loss is based on a novel probabilistic path finding view of dynamic time warping (DTW) that contains the following three key features: (i) the local path routing decisions are contrastive and differentiable, (ii) pairwise distances are cast as probabilities that are contrastive as well, and (iii) our formulation naturally admits a global cycle consistency loss that verifies correspondences. For evaluation, we consider the tasks of fine-grained action classification, few shot learning, and video synchronization. We report significant performance increases over previous methods. In addition, we report two applications of our temporal alignment framework, namely 3D pose reconstruction and fine-grained audio/visual retrieval.

We present a framework for automatically reconfiguring images of street scenes by populating, depopulating, or repopulating them with objects such as pedestrians or vehicles. Applications of this method include anonymizing images to enhance privacy, generating data augmentations for perception tasks like autonomous driving, and composing scenes to achieve a certain ambiance, such as empty streets in the early morning. At a technical level, our work has three primary contributions: (1) a method for clearing images of objects,(2) a method for estimating sun direction from a single image, and (3) a way to compose objects in scenes that respects scene geometry and illumination. Each component is learned from data with minimal ground truth annotations, by making creative use of large-numbers of short image bursts of street scenes. We demonstrate convincing results on a range of street scenes and illustrate potential applications.

This paper strives for repetitive activity counting in videos. Different from existing works, which all analyze the visual video content only, we incorporate for the first time the corresponding sound into the repetition counting process. This benefits accuracy in challenging vision conditions such as occlusion, dramatic camera view changes, low resolution, etc. We propose a model that starts with analyzing the sight and sound streams separately. Then an audiovisual temporal stride decision module and a reliability estimation module are introduced to exploit cross-modal temporal interaction. For learning and evaluation, an existing dataset is repurposed and reorganized to allow for repetition counting with sight and sound. We also introduce a variant of this dataset for repetition counting under challenging vision conditions. Experiments demonstrate the benefit of sound, as well as the other introduced modules, for repetition counting. Our sight-only model already outperforms the state-of-the-art by itself, when we add sound, results improve notably, especially under harsh vision conditions.

We present a simple but powerful architecture of convolutional neural network, which has a VGG-like inference-time body composed of nothing but a stack of 3x3 convolution and ReLU, while the training-time model has a multi-branch topology. Such decoupling of the training-time and inference-time architecture is realized by a structural re-parameterization technique so that the model is named RepVGG. On ImageNet, RepVGG reaches over 80% top-1 accuracy, which is the first time for a plain model, to the best of our knowledge. On NVIDIA 1080Ti GPU, RepVGG models run 83% faster than ResNet-50 or 101% faster than ResNet-101 with higher accuracy and show favorable accuracy-speed trade-off compared to the state-of-the-art models like EfficientNet and RegNet. The code and trained models are available at https://github.com/megvii-model/RepVGG.

Self-supervised learning has shown great potentials in improving the video representation ability of deep neural networks by getting supervision from the data itself. However, some of the current methods tend to cheat from the background, i.e., the prediction is highly dependent on the video background instead of the motion, making the model vulnerable to background changes. To mitigate the model reliance towards the background, we propose to remove the background impact by adding the background. That is, given a video, we randomly select a static frame and add it to every other frames to construct a distracting video sample. Then we force the model to pull the feature of the distracting video and the feature of the original video closer, so that the model is explicitly restricted to resist the background influence, focusing more on the motion changes. We term our method as Background Erasing (BE). It is worth noting that the implementation of our method is so simple and neat and can be added to most of the SOTA methods without much efforts. Specifically, BE brings 16.4% and 19.1% improvements with MoCo on the severely biased datasets UCF101 and HMDB51, and 14.5% improvement on the less biased dataset Diving48.

Existing rain-removal algorithms often tackle either rain streak removal or raindrop removal, and thus may fail to handle real-world rainy scenes. Besides, the lack of real-world deraining datasets comprising different types of rain and their corresponding rain-free ground-truth also impedes deraining algorithm development. In this paper, we aim to address real-world deraining problems from two aspects. First, we propose a complementary cascaded network architecture, namely CCN, to remove rain streaks and raindrops in a unified framework. Specifically, our CCN removes raindrops and rain streaks in a complementary fashion, i.e., raindrop removal followed by rain streak removal and vice versa, and then fuses the results via an attention based fusion module. Considering significant shape and structure differences between rain streaks and raindrops, it is difficult to manually design a sophisticated network to remove them effectively. Thus, we employ neural architecture search to adaptively find optimal architectures within our specified deraining search space. Second, we present a new real-world rain dataset, namely RainDS, to prosper the development of deraining algorithms in practical scenarios. RainDS consists of rain images in different types and their corresponding rain-free ground-truth, including rain streak only, raindrop only, and both of them. Extensive experimental results on both existing benchmarks and RainDS demonstrate that our method outperforms the state-of-the-art.

Recent development of Under-Display Camera (UDC) systems provides a true bezel-less and notch-free viewing experience on smartphones (and TV, laptops, tablets), while allowing images to be captured from the selfie camera embedded underneath. In a typical UDC system, the microstructure of the semi-transparent organic light-emitting diode (OLED) pixel array attenuates and diffracts the incident light on the camera, resulting in significant image quality degradation. Oftentimes, noise, flare, haze, and blur can be observed in UDC images. In this work, we aim to analyze and tackle the aforementioned degradation problems. We define a physics-based image formation model to better understand the degradation. In addition, we utilize one of the world's first commodity UDC smartphone prototypes to measure the real-world Point Spread Function (PSF) of the UDC system, and provide a model-based data synthesis pipeline to generate realistically degraded images. We specially design a new domain knowledge-enabled Dynamic Skip Connection Network (DISCNet) to restore the UDC images. We demonstrate the effectiveness of our method through extensive experiments on both synthetic and real UDC data. Our physics-based image formation model and proposed DISCNet can provide foundations for further exploration in UDC image restoration, and even for general diffraction artifact removal in a broader sense.

With increasing fields of application for neural networks and the development of neural networks, the ability to explain deep learning models is also becoming increasingly important. Especially, prior to practical applications, it is crucial to analyze a model's inference and the process of generating the results. A common explanation method is Class Activation Mapping(CAM) based method where it is often used to understand the last layer of the convolutional neural networks popular in the field of Computer Vision. In this paper, we propose a novel CAM method named Relevance-weighted Class Activation Mapping(Relevance-CAM) that utilizes Layer-wise Relevance Propagation to obtain the weighting components. This allows the explanation map to be faithful and robust to the shattered gradient problem, a shared problem of the gradient based CAM methods that causes noisy saliency maps for intermediate layers. Therefore, our proposed method can better explain a model by correctly analyzing the intermediate layers as well as the last convolutional layer. In this paper, we visualize how each layer of the popular image processing models extracts class specific features using Relevance-CAM, evaluate the localization ability, and show why the gradient based CAM cannot be used to explain the intermediate layers, proven by experimenting the weighting component. Relevance-CAM outperforms other CAM-based methods in recognition and localization evaluation in layers of any depth. The source code is available at: https://github.com/mongeoroo/Relevance-CAM