CVPR 2021 论文列表

RGB-D salient object detection (SOD) is usually formulated as a problem of classification or regression over two modalities, i.e., RGB and depth. Hence, effective RGB-D feature modeling and multi-modal feature fusion both play a vital role in RGB-D SOD. In this paper, we propose a depth-sensitive RGB feature modeling scheme using the depth-wise geometric prior of salient objects. In principle, the feature modeling scheme is carried out in a depth-sensitive attention module, which leads to the RGB feature enhancement as well as the background distraction reduction by capturing the depth geometry prior. Moreover, to perform effective multi-modal feature fusion, we further present an automatic architecture search approach for RGB-D SOD, which does well in finding out a feasible architecture from our specially designed multi-modal multi-scale search space. Extensive experiments on seven standard benchmarks demonstrate the effectiveness of the proposed approach against the state-of-the-art.

We present a novel method for efficient acquisition of shape and spatially varying reflectance of 3D objects using polarization cues. Unlike previous works that have exploited polarization to estimate material or object appearance under certain constraints (known shape or multiview acquisition), we lift such restrictions by coupling polarization imaging with deep learning to achieve high quality estimate of 3D object shape (surface normals and depth) and SVBRDF using single-view polarization imaging under frontal flash illumination. In addition to acquired polarization images, we provide our deep network with strong novel cues related to shape and reflectance, in the form of a normalized Stokes map and an estimate of diffuse color. We additionally describe modifications to network architecture and training loss which provide further qualitative improvements. We demonstrate our approach to achieve superior results compared to recent works employing deep learning in conjunction with flash illumination.

We introduce the concept of rate-aware deep perceptual preprocessing (DPP) for video encoding. DPP makes a single pass over each input frame in order to enhance its visual quality when the video is to be compressed with any codec at any bitrate. The resulting bitstreams can be decoded and displayed at the client side without any post-processing component. DPP comprises a convolutional neural network that is trained via a composite set of loss functions that incorporates: (i) a perceptual loss based on a trained no reference image quality assessment model, (ii) a reference based fidelity loss expressing L1 and structural similarity aspects, (iii) a motion-based rate loss via block-based transform, quantization and entropy estimates that converts the essential components of standard hybrid video encoder designs into a trainable framework. Extensive testing using multiple quality metrics and AVC, AV1 and VVC encoders shows that DPP+encoder reduces, on average, the bitrate of the corresponding encoder by 11%. This marks the first time a server-side neural processing component achieves such savings over the state-of-the-art in video coding.

Many learning-based approaches have difficulty scaling to unseen data, as the generality of its learned prior is limited to the scale and variations of the training samples. This holds particularly true with 3D learning tasks, given the sparsity of 3D datasets available. We introduce a new learning framework for 3D modeling and reconstruction that greatly improves the generalization ability of a deep generator. Our approach strives to connect the good ends of both learning-based and optimization-based methods. In particular, unlike the common practice that fixes the pre-trained priors at test time, we propose to further optimize the learned prior and latent code according to the input physical measurements after the training. We show that the proposed strategy effectively breaks the barriers constrained by the pre-trained priors and could lead to high-quality adaptation to unseen data. We realize our framework using the implicit surface representation and validate the efficacy of our approach in a variety of challenging tasks that take highly sparse or collapsed observations as input. Experimental results show that our approach compares favorably with the state-of-the-art methods in terms of both generality and accuracy.

Segmenting highly-overlapping objects is challenging, because typically no distinction is made between real object contours and occlusion boundaries. Unlike previous two-stage instance segmentation methods, we model image formation as composition of two overlapping layers, and propose Bilayer Convolutional Network (BCNet), where the top GCN layer detects the occluding objects (occluder) and the bottom GCN layer infers partially occluded instance (occludee). The explicit modeling of occlusion relationship with bilayer structure naturally decouples the boundaries of both the occluding and occluded instances, and considers the interaction between them during mask regression. We validate the efficacy of bilayer decoupling on both one-stage and two-stage object detectors with different backbones and network layer choices. Despite its simplicity, extensive experiments on COCO and KINS show that our occlusion-aware BCNet achieves large and consistent performance gain especially for heavy occlusion cases. Code is available at https://github.com/lkeab/BCNet.

Over the last few years, we have witnessed tremendous progress on many subtasks of autonomous driving including perception, motion forecasting, and motion planning. However, these systems often assume that the car is accurately localized against a high-definition map. In this paper we question this assumption, and investigate the issues that arise in state-of-the-art autonomy stacks under localization error. Based on our observations, we design a system that jointly performs perception, prediction, and localization. Our architecture is able to reuse computation between the three tasks, and is thus able to correct localization errors efficiently. We show experiments on a large-scale autonomy dataset, demonstrating the efficiency and accuracy of our proposed approach.

Estimating homography to align image pairs captured by different sensors or image pairs with large appearance changes is an important and general challenge for many computer vision applications. In contrast to others, we propose a generic solution to pixel-wise align multimodal image pairs by extending the traditional Lucas-Kanade algorithm with networks. The key contribution in our method is how we construct feature maps, named as deep Lucas-Kanade feature map (DLKFM). The learned DLKFM can spontaneously recognize invariant features under various appearance-changing conditions. It also has two nice properties for the Lucas-Kanade algorithm: (1) The template feature map keeps brightness consistency with the input feature map, thus the color difference is very small while they are well-aligned. (2) The Lucas-Kanade objective function built on DLKFM has a smooth landscape around ground truth homography parameters, so the iterative solution of the Lucas-Kanade can easily converge to the ground truth. With those properties, directly updating the Lucas-Kanade algorithm on our feature maps will precisely align image pairs with large appearance changes. We share the dataset, code, and demo video online.

Monitoring treatment response in longitudinal studies plays an important role in clinical practice. Accurately identifying lesions across serial imaging follow-up is the core to the monitoring procedure. Typically this incorporates both image and anatomical considerations. However, matching lesions manually is labor-intensive and time-consuming. In this work, we present deep lesion tracker (DLT), a deep learning approach that uses both appearance- and anatomical-based signals. To incorporate anatomical constraints, we propose an anatomical signal encoder, which prevents lesions being matched with visually similar but spurious regions. In addition, we present a new formulation for Siamese networks that avoids the heavy computational loads of 3D cross-correlation. To present our network with greater varieties of images, we also propose a self-supervised learning strategy to train trackers with unpaired images, overcoming barriers to data collection. To train and evaluate our tracker, we introduce and release the first lesion tracking benchmark, consisting of 3891 lesion pairs from the public DeepLesion database. The proposed method, DLT, locates lesion centers with a mean error distance of 7mm. This is 5% better than a leading registration algorithm while running 14 times faster with whole CT volumes. We demonstrate even greater improvements over detector or similarity-learning alternatives. DLT also generalizes well on an external clinical test set of 100% longitudinal studies, achieving 88% accuracy. Finally, we plug DLT into an automatic tumor monitoring workflow where it leads to an accuracy of 85% in assessing lesion treatment responses, which is only 0.46% lower than the accuracy of manual inputs.

Learning-based video compression has achieved substantial progress during recent years. The most influential approaches adopt deep neural networks (DNNs) to remove spatial and temporal redundancies by finding the appropriate lower-dimensional representations of frames in the video. We propose a novel DNN based framework that predicts and compresses video sequences in the latent vector space. The proposed method first learns the efficient lower-dimensional latent space representation of each video frame and then performs inter-frame prediction in that latent domain. The proposed latent domain compression of individual frames is obtained by a deep autoencoder trained with a generative adversarial network (GAN). To exploit the temporal correlation within the video frame sequence, we employ a convolutional long short-term memory (ConvLSTM) network to predict the latent vector representation of the future frame. We demonstrate our method with two applications; video compression and abnormal event detection that share the identical latent frame prediction network. The proposed method exhibits superior or competitive performance compared to the state-of-the-art algorithms specifically designed for either video compression or anomaly detection.

Deep implicit functions (DIFs), as a kind of 3D shape representation, are becoming more and more popular in the 3D vision community due to their compactness and strong representation power. However, unlike polygon mesh-based templates, it remains a challenge to reason dense correspondences or other semantic relationships across shapes represented by DIFs, which limits its applications in texture transfer, shape analysis and so on. To overcome this limitation and also make DIFs more interpretable, we propose Deep Implicit Templates, a new 3D shape representation that supports explicit correspondence reasoning in deep implicit representations. Our key idea is to formulate DIFs as conditional deformations of a template implicit function. To this end, we propose Spatial Warping LSTM, which decomposes the conditional spatial transformation into multiple point-wise transformations and guarantees generalization capability. Moreover, the training loss is carefully designed in order to achieve high reconstruction accuracy while learning a plausible template with accurate correspondences in an unsupervised manner. Experiments show that our method can not only learn a common implicit template for a collection of shapes, but also establish dense correspondences across all the shapes simultaneously without any supervision.

Point set is a flexible and lightweight representation widely used for 3D deep learning. However, their discrete nature prevents them from representing continuous and fine geometry, posing a major issue for learning-based shape generation. In this work, we turn the discrete point sets into smooth surfaces by introducing the well-known implicit moving least-squares (IMLS) surface formulation, which naturally defines locally implicit functions on point sets. We incorporate IMLS surface generation into deep neural networks for inheriting both the flexibility of point sets and the high quality of implicit surfaces. Our IMLSNet predicts an octree structure as a scaffold for generating MLS points where needed and characterizes shape geometry with learned local priors. Furthermore, our implicit function evaluation is independent of the neural network once the MLS points are predicted, thus enabling fast runtime evaluation. Our experiments on 3D object reconstruction demonstrate that IMLSNets outperform state-of-the-art learning-based methods in terms of reconstruction quality and computational efficiency. Extensive ablation tests also validate our network design and loss functions.

In this paper, we propose HESIC, an end-to-end trainable deep network for stereo image compression (SIC). To fully explore the mutual information across two stereo images, we use a deep regression model to estimate the homography matrix, i.e., H matrix. Then, the left image is spatially transformed by the H matrix, and only the residual information between the left and right images is encoded to save bit-rates. A two-branch auto-encoder architecture is adopted in HESIC, corresponding to the left and right images, respectively. For entropy coding, we propose two conditional stereo entropy models, i.e., Gaussian mixture model (GMM) based and context based entropy models, to fully explore the correlation between the two images to reduce the coding bit-rates. In decoding, a cross quality enhancement module is proposed to enhance the image quality based on inverse H matrix. Experimental results show that our HESIC outperforms state-of-the-art SIC methods on InStereo2K and KITTI datasets both quantitatively and qualitatively.

Recently, deep learning based methods have demonstrated promising results on the graph matching problem, by relying on the descriptive capability of deep features extracted on graph nodes. However, one main limitation with existing deep graph matching (DGM) methods lies in their ignorance of explicit constraint of graph structures, which may lead the model to be trapped into local minimum in training. In this paper, we propose to explicitly formulate pairwise graph structures as a quadratic constraint incorporated into the DGM framework. The quadratic constraint minimizes the pairwise structural discrepancy between graphs, which can reduce the ambiguities brought by only using the extracted CNN features. Moreover, we present a differentiable implementation to the quadratic constrained-optimization such that it is compatible with the unconstrained deep learning optimizer. To give more precise and proper supervision, a well-designed false matching loss against class imbalance is proposed, which can better penalize the false negatives and false positives with less overfitting. Exhaustive experiments demonstrate that our method achieves competitive performance on real-world datasets. The code is available at: https://github.com/Zerg-Overmind/QC-DGM.

Pan-sharpening is an important technique for remote sensing imaging systems to obtain high resolution multispectral images. Recently, deep learning has become the most popular tool for pan-sharpening. This paper develops a model-based deep pan-sharpening approach. Specifically, two optimization problems regularized by the deep prior are formulated, and they are separately responsible for the generative models for panchromatic images and low resolution multispectral images. Then, the two problems are solved by a gradient projection algorithm, and the iterative steps are generalized into two network blocks. By alternatively stacking the two blocks, a novel network, called gradient projection based pan-sharpening neural network, is constructed. The experimental results on different kinds of satellite datasets demonstrate that the new network outperforms state-of-the-art methods both visually and quantitatively. The codes are available at https://github.com/xsxjtu/GPPNN.

In coded aperture snapshot spectral imaging (CASSI) system, the real-world hyperspectral image (HSI) can be reconstructed from the captured compressive image in a snapshot. Model-based HSI reconstruction methods employed hand-crafted priors to solve the reconstruction problem, but most of which achieved limited success due to the poor representation capability of these hand-crafted priors. Deep learning based methods learning the mappings between the compressive images and the HSIs directly achieved much better results. Yet, it is nontrivial to design a powerful deep network heuristically for achieving satisfied results. In this paper, we propose a novel HSI reconstruction method based on the Maximum a Posterior (MAP) estimation framework using learned Gaussian Scale Mixture (GSM) prior. Different from existing GSM models using hand-crafted scale priors (e.g., the Jeffrey's prior), we propose to learn the scale prior through a deep convolutional neural network (DCNN). Furthermore, we also propose to estimate the local means of the GSM models by the DCNN. All the parameters of the MAP estimation algorithm and the DCNN parameters are jointly optimized through end-to-end training. Extensive experimental results on both synthetic and real datasets demonstrate that the proposed method outperforms existing state-of-the-art methods. The code is available at https://see.xidian.edu.cn/faculty/wsdong/Projects/DGSM-SCI.htm.

Multi-frame human pose estimation in complicated situations is challenging. Although state-of-the-art human joints detectors have demonstrated remarkable results for static images, their performances come short when we apply these models to video sequences. Prevalent shortcomings include the failure to handle motion blur, video defocus, or pose occlusions, arising from the inability in capturing the temporal dependency among video frames. On the other hand, directly employing conventional recurrent neural networks incurs empirical difficulties in modeling spatial contexts, especially for dealing with pose occlusions. In this paper, we propose a novel multi-frame human pose estimation framework, leveraging abundant temporal cues between video frames to facilitate keypoint detection. Three modular components are designed in our framework. A Pose Temporal Merger encodes keypoint spatiotemporal context to generate effective searching scopes while a Pose Residual Fusion module computes weighted pose residuals in dual directions. These are then processed via our Pose Correction Network for efficient refining of pose estimations. Our method ranks No.1 in the Multi-frame Person Pose Estimation Challenge on the large-scale benchmark datasets PoseTrack2017 and PoseTrack2018. We have released our code, hoping to inspire future research.

We introduce a neural network-based method to denoise pairs of images taken in quick succession in low-light environments, with and without a flash. Our goal is to produce a high-quality rendering of the scene that preserves the color and mood from the ambient illumination of the noisy no-flash image, while recovering surface texture and detail revealed by the flash. Our network outputs a gain map and a field of kernels, the latter obtained by linearly mixing elements of a per-image low-rank kernel basis. We first apply the kernel field to the no-flash image, and then multiply the result with the gain map to create the final output. We show our network effectively learns to produce high-quality images by combining a smoothed out estimate of the scene's ambient appearance from the no-flash image, with high-frequency albedo details extracted from the flash input. Our experiments show significant improvements over alternative captures without a flash, and baseline denoisers that use flash no-flash pairs. In particular, our method produces images that are both noise-free and contain accurate ambient colors without the sharp shadows or strong specular highlights visible in the flash image.

Inspired by the great success of deep neural networks (DNNs), many unfolding methods have been proposed to integrate traditional image modeling techniques, such as dictionary learning (DicL) and sparse coding, into DNNs for image restoration. However, the performance of such methods remains limited for several reasons. First, the unfolded architectures do not strictly follow the image representation model of DicL and lose the desired physical meaning. Second, handcrafted priors are still used in most unfolding methods without effectively utilizing the learning capability of DNNs. Third, a universal dictionary is learned to represent all images, reducing the model representation flexibility. We propose a novel framework of deep convolutional dictionary learning (DCDicL), which follows the representation model of DicL strictly, learns the priors for both representation coefficients and the dictionaries, and can adaptively adjust the dictionary for each input image based on its content. The effectiveness of our DCDicL method is validated on the image denoising problem. DCDicL demonstrates leading denoising performance in terms of both quantitative metrics (e.g., PSNR, SSIM) and visual quality. In particular, it can reproduce the subtle image structures and textures, which are hard to recover by many existing denoising DNNs. The code is available at: https://github.com/natezhenghy/DCDicL_denoising.

In this paper, we propose a deep compositional metric learning (DCML) framework for effective and generalizable similarity measurement between images. Conventional deep metric learning methods minimize a discriminative loss to enlarge interclass distances while suppressing intraclass variations, which might lead to inferior generalization performance since samples even from the same class may present diverse characteristics. This motivates the adoption of the ensemble technique to learn a number of sub-embeddings using different and diverse subtasks. However, most subtasks impose weaker or contradictory constraints, which essentially sacrifices the discrimination ability of each sub-embedding to improve the generalization ability of their combination. To achieve a better generalization ability without compromising, we propose to separate the sub-embeddings from direct supervisions from the subtasks and apply the losses on different composites of the sub-embeddings. We employ a set of learnable compositors to combine the sub-embeddings and use a self-reinforced loss to train the compositors, which serve as relays to distribute the diverse training signals to avoid destroying the discrimination ability. Experimental results on the CUB-200-2011, Cars196, and Stanford Online Products datasets demonstrate the superior performance of our framework.