ECCV 2020 论文列表

Although two-stage object detectors have continuously advanced the state-of-the-art performance in recent years, the training process itself is far from crystal. In this work, we first point out the inconsistency problem between the fixed network settings and the dynamic training procedure, which greatly affects the performance. For example, the fixed label assignment strategy and regression loss function cannot fit the distribution change of proposals and thus are harmful to training high quality detectors. Consequently, we propose Dynamic R-CNN to adjust the label assignment criteria (IoU threshold) and the shape of regression loss function (parameters of SmoothL1 Loss) automatically based on the statistics of proposals during training. This dynamic design makes better use of the training samples and pushes the detector to fit more high quality samples. Specifically, our method improves upon ResNet-50-FPN baseline with 1.9% AP and 5.5% AP$_{90}$ on the MS COCO dataset with no extra overhead. Codes and models are available at https://github.com/hkzhang95/DynamicRCNN."

Imaging in low light is difficult because the number of photons arriving at the sensor is low. Imaging dynamic scenes in low-light environments is even more difficult because as the scene moves, pixels in adjacent frames need to be aligned before they can be denoised. Conventional CMOS image sensors (CIS) are at a particular disadvantage in dynamic low-light settings because the exposure cannot be too short lest the read noise overwhelms the signal. We propose a solution using Quanta Image Sensors (QIS) and present a new image reconstruction algorithm. QIS are single-photon image sensors with photon counting capabilities. Studies over the past decade have confirmed the effectiveness of QIS for low-light imaging but reconstruction algorithms for dynamic scenes in low light remain an open problem. We fill the gap by proposing a student-teacher training protocol that transfers knowledge from a motion teacher and a denoising teacher to a student network. We show that dynamic scenes can be reconstructed from a burst of frames at a photon level of 1 photon per pixel per frame. Experimental results confirm the advantages of the proposed method compared to existing methods."

Replacing normal convolutions with group convolutions can significantly increase the computational efficiency of modern deep convolutional networks, which has been widely adopted in compact network architecture designs. However, existing group convolutions undermine the original network structures by cutting off some connections permanently resulting in significant accuracy degradation. In this paper, we propose dynamic group convolution (DGC) that adaptively selects which part of input channels to be connected within each group for individual samples on the fly. Specifically, we equip each group with a small feature selector to automatically select the most important input channels conditioned on the input images. Multiple groups can adaptively capture abundant and complementary visual/semantic features for each input image. The DGC preserves the original network structure and has similar computational efficiency as the conventional group convolution simultaneously. Extensive experiments on multiple image classification benchmarks including CIFAR-10, CIFAR-100 and ImageNet demonstrate its superiority over the existing group convolution techniques and dynamic execution methods. The code is available at https://github.com/zhuogege1943/dgc."

Visible-infrared person re-identification (VI-ReID) is a challenging cross-modality pedestrian retrieval problem. Due to the large intra-class variations and cross-modality discrepancy with large amount of sample noise, it is difficult to learn discriminative part features. Existing VI-ReID methods instead tend to learn global representations, which have limited discriminability and weak robustness to noisy images. In this paper, we propose a novel dynamic dual-attentive aggregation (DDAG) learning method by mining both intra-modality part-level and cross-modality graph-level contextual cues for VI-ReID. We propose an intra-modality weighted-part attention module to extract discriminative part-aggregated features, by imposing the domain knowledge on the part relationship mining. To enhance robustness against noisy samples, we introduce cross-modality graph structured attention to reinforce the representation with the contextual relations across the two modalities. We also develop a parameter-free dynamic dual aggregation learning strategy to adaptively integrate the two components in a progressive joint training manner. Extensive experiments demonstrate that DDAG outperforms the state-of-the-art methods under various settings."

Computational stereo has reached a high level of accuracy, but degrades in the presence of occlusions, repeated textures, and correspondence errors along edges. We present a novel approach based on neural networks for depth estimation that combines stereo from dual cameras with stereo from a dual-pixel sensor, which is increasingly common on consumer cameras. Our network uses a novel architecture to fuse these two sources of information and can overcome the above-mentioned limitations of pure binocular stereo matching. Our method provides a dense depth map with sharp edges, which is crucial for computational photography applications like synthethic shallow-depth-of-field or 3D Photos. Additionally, we avoid the inherent ambiguity due to the aperture problem in stereo cameras by designing the stereo baseline to be orthogonal to the dual-pixel baseline. We present experiments and comparisons with state-of-the-art approaches that show that our method offers a substantial improvement over previous works."

In this paper, we tackle an open research question in transfer learning, which is selecting a model initialization to achieve high performance on a new task, given several pre-trained models. We propose a new highly efficient and accurate approach based on duality diagram similarity (DDS) between deep neural networks (DNNs). DDS is a generic framework to represent and compare data of different feature dimensions. We validate our approach on the Taskonomy dataset by measuring the correspondence between actual transfer learning performance rankings on 17 taskonomy tasks and predicted rankings. Computing DDS based ranking for $17 imes17$ transfers requires less than 2 minutes and shows a high correlation ($0.86$) with actual transfer learning rankings, outperforming state-of-the-art methods by a large margin ($10\%$) on the Taskonomy benchmark. We also demonstrate the robustness of our model selection approach to a new task, namely Pascal VOC semantic segmentation. Additionally, we show that our method can be applied to select the best layer locations within a DNN for transfer learning on 2D, 3D and semantic tasks on NYUv2 and Pascal VOC datasets."

Due to the complex underwater environment, underwater imaging often encounters some problems such as blur, scale variation, color shift, and texture distortion. Generic detection algorithms can not work well when we use them directly in the underwater scene. To address these problems, we propose an underwater detection framework with feature enhancement and anchor refinement. It has a composite connection backbone to boost the feature representation and introduces a receptive field augmentation module to exploit multi-scale contextual features. The developed underwater object detection framework also provides a prediction refinement scheme according to six prediction layers, it can refine multi-scale features to better align with anchors by learning from offsets, which solve the problem of sample imbalance to a certain extent. We also construct a new underwater detection dataset, denoted as UWD, which has more than 10,000 train-val and test underwater images. The extensive experiments on PASCAL VOC and UWD demonstrate the favorable performance of the proposed underwater detection framework against the states-of-the-arts methods in terms of accuracy and robustness. Source code and models are available at: https://github.com/Peterchen111/FERNet."

Recent advances on unsupervised domain adaptation (UDA) rely on adversarial learning to disentangle the explanatory and transferable features for domain adaptation. However, there are two issues with the existing methods. First, the discriminability of the latent space cannot be fully guaranteed without considering the class-aware information in the target domain. Second, samples from the source and target domains alone are not sufficient for domain-invariant feature extracting in the latent space. In order to alleviate the above issues, we propose a dual mixup regularized learning (DMRL) method for UDA, which not only guides the classifier in enhancing consistent predictions in-between samples, but also enriches the intrinsic structures of the latent space. The DMRL jointly conducts category and domain mixup regularizations on pixel level to improve the effectiveness of models. A series of empirical studies on four domain adaptation benchmarks demonstrate that our approach can achieve the state-of-the-art."

We aim to recover the dense 3D surface of the hand from depth maps and propose a network that can predict mesh vertices, transformation matrices for every joint and joint coordinates in a single forward pass. Use fully convolutional architectures, we first map depth image features to the mesh grid and then regress the mesh coordinates into real world 3D coordinates. The final mesh is found by sampling from the mesh grid refit in closed-form based on an articulated template mesh.

Real-world image noise removal is a long-standing yet very challenging task in computer vision. The success of deep neural network in denoising stimulates the research of noise generation, aiming at synthesizing more pairs of clean-noisy images to facilitate the training of deep. In this work, we propose a novel unified framework to simultaneously deal with the noise removal and noise generation tasks. Instead of only inferring the posterior distribution of the latent clean image conditioned on the observed noisy image in traditional MAP framework, it learns the joint distribution of the clean-noisy image pairs. Specifically, we approximate the joint distribution with two different factorized forms, which can be formulated as a denoiser mapping the noisy image to the clean one and a generator mapping the clean to the noisy one. The learned joint distribution implicitly contains all the information between the noisy and clean images, avoiding the necessity of manually designing the image priors and noise assumptions as traditional. Besides, the performance of our denoiser can be further improved by augmenting the original training dataset with the learned generator. Moreover, we propose two metrics to assess the quality of the generated noisy image, for which, to the best of our knowledge, such metrics are first proposed along this research line. Extensive experiments have conducted to demonstrate the superiority of our method over the state-of-the-arts in terms of both the real noise removal and noise generation tasks."

Active learning, reducing the cost and workload of annotations, attracts increasing attentions from the community. Current active learning approaches commonly adopted uncertainty-based acquisition functions for the data selection due to their effectiveness. However, data selection based on uncertainty suffers from the overlapping problem, i.e., the top-$K$ samples ranked by the uncertainty are similar. In this paper, we investigate the overlapping problem of recent uncertainty-based approaches and propose to alleviate the issue by taking representativeness into consideration. In particular, we propose a dual adversarial network, namely DAAL, for this purpose. Different from previous hybrid active learning methods requiring multi-stage data selections i.e., step-by-step evaluating the uncertainty and representativeness using different acquisition functions, our DAAL learns to select the most uncertain and representative data points in one-stage. Extensive experiments conducted on three publicly available datasets, i.e., CIFAR10/100 and Cityscapes, demonstrate the effectiveness of our method---a new state-of-the-art accuracy is achieved."

Conventional unsupervised domain adaptation (UDA) studies the knowledge transfer between a limited number of domains. This neglects the more practical scenario where data are distributed in numerous different domains in the real world. The domain similarity between those domains is critical for domain adaptation performance. To describe and learn relations between different domains, we propose a novel Domain2Vec model to provide vectorial representations of visual domains based on joint learning of feature disentanglement and Gram matrix. To evaluate the effectiveness of our Domain2Vec model, we create two large-scale cross-domain benchmarks. The first one is TinyDA, which contains 54 domains and about one million MNIST-style images. The second benchmark is DomainBank, which is collected from 56 existing vision datasets. We demonstrate that our embedding is capable of predicting domain similarities that match our intuition about visual relations between different domains. Extensive experiments are conducted to demonstrate the power of our new datasets in benchmarking state-of-the-art multi-source domain adaptation methods, as well as the advantage of our proposed model. "

State-of-the-art stereo matching networks have difficulties in generalizing to new unseen environments due to significant domain differences, such as color, illumination, contrast, and texture. In this paper, we aim at designing a domain-invariant stereo matching network (DSMNet) that generalizes well to unseen scenes. To achieve this goal, we propose i) a novel ``""domain normalization"" approach that regularizes the distribution of learned representations to allow them to be invariant to domain differences, and ii) a trainable non-local graph-based filter for extracting robust structural and geometric representations that can further enhance domain-invariant generalizations. When trained on synthetic data and generalized to real test sets, our model performs significantly better than all state-of-the-art models. It even outperforms some deep learning models (e.g. DispNet, Content-CNN and MC-CNN) fine-tuned with test-domain data."

Style transfer generates an image whose content comes from one image and style from the other. Image-to-image translation approaches with disentangled representations have been shown effective for style transfer between two image categories. However, previous methods often assume a shared domain-invariant content space, which could compromise the content representation power. For addressing this issue, this paper leverages domain-specific mappings for remapping latent features in the shared content space to domain-specific content spaces. This way, images can be encoded more properly for style transfer. Experiments show that the proposed method outperforms previous style transfer methods, particularly on challenging scenarios that would require semantic correspondences between images. Code and results are available at https://github.com/acht7111020/DSMAP."

We propose a novel framework for domain adaptation in semantic segmentation with image-level weak labels in the target domain. The weak labels may be obtained based on a model prediction for unsupervised domain adaptation (UDA), or from a human oracle in a new weakly-supervised domain adaptation (WDA) paradigm for semantic segmentation. Using weak labels is both practical and useful, since (i) collecting image-level target annotations is comparably cheap in WDA and incurs no cost in UDA, and (ii) it opens the opportunity for category-wise domain alignment. Our framework uses weak labels to enable the interplay between feature alignment and pseudo-labeling, improving both in the process of domain adaptation. Specifically, we develop a weak-label classification module to enforce the network to attend to certain categories, and then use such training signals to guide the proposed category-wise alignment method. In experiments, we show considerable improvements with respect to the existing state-of-the-arts in UDA and present a new benchmark in the WDA setting."

Conventional object detection models inevitably encounter a performance drop as the domain disparity exists. Unsupervised domain adaptive object detection is proposed recently to reduce the disparity between domains, where the source domain is label-rich while the target domain is label-agnostic. The existing models follow a parameter shared siamese structure for adversarial domain alignment, which, however, easily leads to the collapse and out-of-control risk of the source domain and brings negative impact to feature adaption. The main reason is that the labeling unfairness (asymmetry) between source and target makes the parameter sharing mechanism unable to adapt. Therefore, in order to avoid the source domain collapse risk caused by parameter sharing, we propose an asymmetric tri-way Faster-RCNN (ATF) for domain adaptive object detection. Our ATF model has two distinct merits: 1) A ancillary net supervised by source label is deployed to learn ancillary target features and simultaneously preserve the discrimination of source domain, which enhances the structural discrimination (object classification vs. bounding box regression) of domain alignment. 2) The asymmetric structure consisting of a chief net and an independent ancillary net essentially overcomes the parameter sharing aroused source risk collapse. The adaption safety of the proposed ATF detector is guaranteed. Extensive experiments on a number of datasets, including Cityscapes, Foggy-cityscapes, KITTI, Sim10k, Pascal VOC, Clipart and Watercolor, demonstrate the SOTA performance of our method."

Deep networks devour millions of precisely annotated images to build their complex and powerful representations. Unfortunately, tasks like autonomous driving have virtually no real-world training data. Repeatedly crashing a car into a tree is simply too expensive. The commonly prescribed solution is simple: learn a representation in simulation and transfer it to the real world. However, this transfer is challenging since simulated and real-world visual experiences vary dramatically. Our core observation is that for certain tasks, such as image recognition, datasets are plentiful. They exist in any interesting domain, simulated or real, and are easy to label and extend. We use these recognition datasets to link up a source and target domain to transfer models between them in a task distillation framework. Our method can successfully transfer navigation policies between drastically different simulators: ViZDoom, SuperTuxKart, and CARLA. Furthermore, it shows promising results on standard domain adaptation benchmarks."

Structure extraction from document images has been a long-standing research topic due to its high impact on a wide range of practical applications. In this paper, we share our findings on employing a hierarchical semantic segmentation network for this task of structure extraction. We propose a prior based deep hierarchical CNN network architecture that enables document structure extraction using very high resolution(1800 x 1000) images. We divide the document image into overlapping horizontal strips such that the network segments a strip and uses its prediction mask as prior for predicting the segmentation of the subsequent strip. We perform experiments establishing the effectiveness of our strip based network architecture through ablation methods and comparison with low-resolution variations. Further, to demonstrate our network's capabilities, we train it on only one type of documents (Forms) and achieve state-of-the-art results over other general document datasets. We introduce our new human-annotated forms dataset and show that our method significantly outperforms different segmentation baselines on this dataset in extracting hierarchical structures.Our method is currently being used in Adobe's AEM Forms for automated conversion of paper and PDF forms to modern HTML based forms."

The 2D virtual try-on task has recently attracted a great interest from the research community, for its direct potential applications in online shopping as well as for its inherent and non-addressed scientific challenges. This task requires fitting an in-shop cloth image on the image of a person, which is highly challenging because it involves cloth warping, image compositing, and synthesizing. Casting virtual try-on into a supervised task faces a difficulty: available datasets are composed of pairs of pictures (cloth, person wearing the cloth). Thus, we have no access to ground-truth when the cloth on the person changes. State-of-the-art models solve this by masking the cloth information on the person with both a human parser and a pose estimator. Then, image synthesis modules are trained to reconstruct the person image from a masked person image and a cloth image. This procedure has several caveats: firstly, human parsers are prone to errors; secondly, it is a costly pre-processing step, which also has to be applied at inference time; finally, it makes the task harder than it is since the mask covers information that should be kept such as hands or accessories. In this paper, we propose a novel student-teacher paradigm where the teacher is trained in the standard way (reconstruction) before guiding the student to focus on the initial task (changing the cloth). The student additionally learns from an adversarial loss, which pushes it to follow the distribution of the real images. Consequently, the student exploits information that is masked to the teacher. A student trained without the adversarial loss would not use this information. Also, getting rid of both human parser and pose estimator at inference time allows obtaining a real-time virtual try-on."