ICCV 2021 Papers

The popular unsupervised video object segmentation methods fuse the RGB frame and optical flow via a two-stream network. However, they cannot handle the distracting noises in each input modality, which may vastly deteriorate the model performance. We propose to establish the correspondence between the input modalities while suppressing the distracting signals via optimal structural matching. Given a video frame, we extract the dense local features from the RGB image and optical flow, and treat them as two complex structured representations. The Wasserstein distance is then employed to compute the global optimal flows to transport the features in one modality to the other, where the magnitude of each flow measures the extent of the alignment between two local features. To plug the structural matching into a two-stream network for end-to-end training, we factorize the input cost matrix into small spatial blocks and design a differentiable long-short Sinkhorn module consisting of a long-distant Sinkhorn layer and a short-distant Sinkhorn layer. We integrate the module into a dedicated two-stream network and dub our model TransportNet. Our experiments show that aligning motion-appearance yields the state-of-the-art results on the popular video object segmentation datasets.

Underexposed image enhancement is of importance in many research domains. In this paper, we take this problem as image feature transformation between the underexposed image and its paired enhanced version, and we propose a deep symmetric network for the issue. Our symmetric network adapts invertible neural networks (INN) for bidirectional feature learning between images, and to ensure the mutual propagation invertible we specifically construct two pairs of encoder-decoder with the same pretrained parameters. This invertible mechanism with bidirectional feature transformations enable us to both avoid colour bias and recover the content effectively for image enhancement. In addition, we propose a new recurrent residual-attention module (RRAM), where the recurrent learning network is designed to gradually perform the desired colour adjustments. Ablation experiments are executed to show the role of each component of our new architecture. We conduct a large number of experiments on two datasets to demonstrate that our method achieves the state-of-the-art effect in underexposed image enhancement. Code is available at https://www.shaopinglu.net/proj-iccv21/ImageEnhancement.html

Time-to-event analysis is an important statistical tool for allocating clinical resources such as ICU beds. However, classical techniques like the Cox model cannot directly incorporate images due to their high dimensionality. We propose a deep learning approach that naturally incorporates multiple, time-dependent imaging studies as well as non-imaging data into time-to-event analysis. Our techniques are benchmarked on a clinical dataset of 1,894 COVID-19 patients, and show that image sequences significantly improve predictions. For example, classical time-to-event methods produce a concordance error of around 30-40% for predicting hospital admission, while our error is 25% without images and 20% with multiple X-rays included. Ablation studies suggest that our models are not learning spurious features such as scanner artifacts. While our focus and evaluation is on COVID-19, the methods we develop are broadly applicable.

Effectively structuring deep knowledge plays a pivotal role in transfer from teacher to student, especially in semantic vision tasks. In this paper, we present a simple knowledge structure to exploit and encode information inside the detection system to facilitate detector knowledge distillation. Specifically, aiming at solving the feature imbalance problem while further excavating the missing relation inside semantic instances, we design a graph whose nodes correspond to instance proposal-level features and edges represent the relation between nodes. To further refine this graph, we design an adaptive background loss weight to reduce node noise and background samples mining to prune trivial edges. We transfer the entire graph as encoded knowledge representation from teacher to student, capturing local and global information simultaneously. We achieve new state-of-the-art results on the challenging COCO object detection task with diverse student-teacher pairs on both one- and two-stage detectors. We also experiment with instance segmentation to demonstrate robustness of our method. It is notable that distilled Faster R-CNN with ResNet18-FPN and ResNet50-FPN yields 38.68 and 41.82 Box AP respectively on the COCO benchmark, Faster R-CNN with ResNet101-FPN significantly achieves 43.38 AP, which outperforms ResNet152-FPN teacher about 0.7 AP. Code: https://github.com/dvlab-research/Dsig.

We propose a deep reparametrization of the maximum a posteriori formulation commonly employed in multi-frame image restoration tasks. Our approach is derived by introducing a learned error metric and a latent representation of the target image, which transforms the MAP objective to a deep feature space. The deep reparametrization allows us to directly model the image formation process in the latent space, and to integrate learned image priors into the prediction. Our approach thereby leverages the advantages of deep learning, while also benefiting from the principled multi-frame fusion provided by the classical MAP formulation. We validate our approach through comprehensive experiments on burst denoising and burst super-resolution datasets. Our approach sets a new state-of-the-art for both tasks, demonstrating the generality and effectiveness of the proposed formulation.

This paper presents a deep relational metric learning (DRML) framework for image clustering and retrieval. Most existing deep metric learning methods learn an embedding space with a general objective of increasing interclass distances and decreasing intraclass distances. However, the conventional losses of metric learning usually suppress intraclass variations which might be helpful to identify samples of unseen classes. To address this problem, we propose to adaptively learn an ensemble of features that characterizes an image from different aspects to model both interclass and intraclass distributions. We further employ a relational module to capture the correlations among each feature in the ensemble and construct a graph to represent an image. We then perform relational inference on the graph to integrate the ensemble and obtain a relation-aware embedding to measure the similarities. Extensive experiments on the widely-used CUB-200-2011, Cars196, and Stanford Online Products datasets demonstrate that our framework improves existing deep metric learning methods and achieves very competitive results.

Existing deep methods produce highly accurate 3D reconstructions in stereo and multiview stereo settings, i.e., when cameras are both internally and externally calibrated. Nevertheless, the challenge of simultaneous recovery of camera poses and 3D scene structure in multiview settings with deep networks is still outstanding. Inspired by projective factorization for Structure from Motion (SFM) and by deep matrix completion techniques, we propose a neural network architecture that, given a set of point tracks in multiple images of a static scene, recovers both the camera parameters and a (sparse) scene structure by minimizing an unsupervised reprojection loss. Our network architecture is designed to respect the structure of the problem: the sought output is equivariant to permutations of both cameras and scene points. Notably, our method does not require initialization of camera parameters or 3D point locations. We test our architecture in two setups: (1) single scene reconstruction and (2) learning from multiple scenes. Our experiments, conducted on a variety of datasets in both internally calibrated and uncalibrated settings, indicate that our method accurately recovers pose and structure, on par with classical state of the art methods. Additionally, we show that a pre-trained network can be used to reconstruct novel scenes using inexpensive fine-tuning with no loss of accuracy.

Classical close-set semantic segmentation networks have limited ability to detect out-of-distribution (OOD) objects, which is important for safety-critical applications such as autonomous driving. Incrementally learning these OOD objects with few annotations is an ideal way to enlarge the knowledge base of the deep learning models. In this paper, we propose an open world semantic segmentation system that includes two modules: (1) an open-set semantic segmentation module to detect both in-distribution and OOD objects. (2) an incremental few-shot learning module to gradually incorporate those OOD objects into its existing knowledge base. This open world semantic segmentation system behaves like a human being, which is able to identify OOD objects and gradually learn them with corresponding supervision. We adopt the Deep Metric Learning Network (DMLNet) with contrastive clustering to implement open-set semantic segmentation. Compared to other open-set semantic segmentation methods, our DMLNet achieves state-of-the-art performance on three challenging open-set semantic segmentation datasets without using additional data or generative models. On this basis, two incremental few-shot learning methods are further proposed to progressively improve the DMLNet with the annotations of OOD objects.

Conventional techniques to establish dense correspondences across visually or semantically similar images focused on designing a task-specific matching prior, which is difficult to model in general. To overcome this, recent learning-based methods have attempted to learn a good matching prior within a model itself on large training data. The performance improvement was apparent, but the need for sufficient training data and intensive learning hinders their applicability. Moreover, using the fixed model at test time does not account for the fact that a pair of images may require their own prior, thus providing limited performance and poor generalization to unseen images. In this paper, we show that an image pair-specific prior can be captured by solely optimizing the untrained matching networks on an input pair of images. Tailored for such test-time optimization for dense correspondence, we present a residual matching network and a confidence-aware contrastive loss to guarantee a meaningful convergence. Experiments demonstrate that our framework, dubbed Deep Matching Prior (DMP), is competitive, or even outperforms, against the latest learning-based methods on several benchmarks for geometric matching and semantic matching, even though it requires neither large training data nor intensive learning. With the networks pre-trained, DMP attains state-of-the-art performance on all benchmarks.

Deep neural representations of 3D shapes as implicit functions have been shown to produce high fidelity models surpassing the resolution-memory trade-off faced by the explicit representations using meshes and point clouds. However, most such approaches focus on representing closed shapes. Unsigned distance function (UDF) based approaches have been proposed recently as a promising alternative to represent both open and closed shapes. However, since the gradients of UDFs vanish on the surface, it is challenging to estimate local (differential) geometric properties like the normals and tangent planes which are needed for many downstream applications in vision and graphics. There are additional challenges in computing these properties efficiently with a low-memory footprint. This paper presents a novel approach that models such surfaces using a new class of implicit representations called the closest surface-point CSP representation. We show that CSP allows us to represent complex surfaces of any topology (open or closed) with high fidelity. It also allows for accurate and efficient computation of local geometric properties. We further demonstrate that it leads to efficient implementation of downstream algorithms like sphere-tracing for rendering the 3D surface as well as to create explicit mesh-based representations. Extensive experimental evaluation on the ShapeNet dataset validate the above contributions with results surpassing the state-of-the-art. Code and data are available at https://sites.google.com/view/cspnet

We present a deep learning pipeline that leverages network self-prior to recover a full 3D model consisting of both a triangular mesh and a texture map from the colored 3D point cloud. Different from previous methods either exploiting 2D self-prior for image editing or 3D self-prior for pure surface reconstruction, we propose to exploit a novel hybrid 2D-3D self-prior in deep neural networks to significantly improve the geometry quality and produce a high-resolution texture map, which is typically missing from the output of commodity-level 3D scanners. In particular, we first generate an initial mesh using a 3D convolutional neural network with 3D self-prior, and then encode both 3D information and color information in the 2D UV atlas, which is further refined by 2D convolutional neural networks with the self-prior. In this way, both 2D and 3D self-priors are utilized for the mesh and texture recovery. Experiments show that, without the need of any additional training data, our method recovers the 3D textured mesh model of high quality from sparse input, and outperforms the state-of-the-art methods in terms of both the geometry and texture quality.

Point cloud registration is the task of estimating the rigid transformation that aligns a pair of point cloud fragments. We present an efficient and robust framework for pairwise registration of real-world 3D scans, leveraging Hough voting in the 6D transformation parameter space. First, deep geometric features are extracted from a point cloud pair to compute putative correspondences. We then construct a set of triplets of correspondences to cast votes on the 6D Hough space, which represents the transformation parameters in the form of sparse tensors. Next, a fully convolutional refinement module is applied to refine the noisy votes. Finally, we identify the consensus among the correspondences from the Hough space, which we use to predict our final transformation parameters. Our method outperforms state-of-the-art methods on the 3DMatch and 3DLoMatch benchmarks while achieving comparable performance on the KITTI odometry dataset. We further demonstrate the generalizability of our approach by setting a new state-of-the-art on the ICL-NUIM dataset, where we integrate our module into a multi-way registration pipeline.

Existing halftoning algorithms usually drop colors and fine details when dithering color images with binary dot patterns, which makes it extremely difficult to recover the original information. To dispense the recovery trouble in future, we propose a novel halftoning technique that converts a color image into binary halftone with full restorability to the original version. The key idea is to implicitly embed those previously dropped information into the halftone patterns. So, the halftone pattern not only serves to reproduce the image tone, maintain the blue-noise randomness, but also represents the color information and fine details. To this end, we exploit two collaborative convolutional neural networks (CNNs) to learn the dithering scheme, under a non-trivial self-supervision formulation. To tackle the flatness degradation issue of CNNs, we propose a novel noise incentive block (NIB) that can serve as a generic CNN plug-in for performance promotion. At last, we tailor a guiding-aware training scheme that secures the convergence direction as regulated. We evaluate the invertible halftones in multiple aspects, which evidences the effectiveness of our method.

Deep neural networks for automatic image colorization often suffer from the color-bleeding artifact, a problematic color spreading near the boundaries between adjacent objects. Such color-bleeding artifacts debase the reality of generated outputs, limiting the applicability of colorization models in practice. Although previous approaches have attempted to address this problem in an automatic manner, they tend to work only in limited cases where a high contrast of gray-scale values are given in an input image. Alternatively, leveraging user interactions would be a promising approach for solving this color-breeding artifacts. In this paper, we propose a novel edge-enhancing network for the regions of interest via simple user scribbles indicating where to enhance. In addition, our method requires a minimal amount of effort from users for their satisfactory enhancement. Experimental results demonstrate that our interactive edge-enhancing approach effectively improves the color-bleeding artifacts compared to the existing baselines across various datasets.

Semi-supervised domain adaptation (SSDA) aims to adapt models trained from a labeled source domain to a different but related target domain, from which unlabeled data and a small set of labeled data are provided. Current methods that treat source and target supervision without distinction overlook their inherent discrepancy, resulting in a source-dominated model that has not effectively use the target supervision. In this paper, we argue that the labeled target data needs to be distinguished for effective SSDA, and propose to explicitly decompose the SSDA task into two sub-tasks: a semi-supervised learning (SSL) task in the target domain and an unsupervised domain adaptation (UDA) task across domains. By doing so, the two sub-tasks can better leverage the corresponding supervision and thus yield very different classifiers. To integrate the strengths of the two classifiers, we apply the well established co-training framework, in which the two classifiers exchange their high confident predictions to iteratively "teach each other" so that both classifiers can excel in the target domain. We call our approach Deep Co-training with Task decomposition (DeCoTa). DeCoTa requires no adversarial training and is easy to implement. Moreover, DeCoTa is well founded on the theoretical condition of when co-training would succeed. As a result, DeCoTa achieves state-of-the-art results on several SSDA datasets, outperforming the prior art by a notable 4% margin on DomainNet. Code is available at https://github.com/LoyoYang/DeCoTa.

Existing video super-resolution (SR) algorithms usually assume that the blur kernels in the degradation process are known and do not model the blur kernels in the restoration. However, this assumption does not hold for blind video SR and usually leads to over-smoothed super-resolved frames. In this paper, we propose an effective blind video SR algorithm based on deep convolutional neural networks (CNNs). Our algorithm first estimates blur kernels from low-resolution (LR) input videos. Then, with the estimated blur kernels, we develop an effective image deconvolution method based on the image formation model of blind video SR to generate intermediate latent frames so that sharp image contents can be restored well. To effectively explore the information from adjacent frames, we estimate the motion fields from LR input videos, extract features from LR videos by a feature extraction network, and warp the extracted features from LR inputs based on the motion fields. Moreover, we develop an effective sharp feature exploration method which first extracts sharp features from restored intermediate latent frames and then uses a transformation operation based on the extracted sharp features and warped features from LR inputs to generate better features for HR video restoration. We formulate the proposed algorithm into an end-to-end trainable framework and show that it performs favorably against state-of-the-art methods.

Image view synthesis has seen great success in reconstructing photorealistic visuals, thanks to deep learning and various novel representations. The next key step in immersive virtual experiences is view synthesis of dynamic scenes. However, several challenges exist due to the lack of high-quality training datasets, and the additional time dimension for videos of dynamic scenes. To address this issue, we introduce a multi-view video dataset, captured with a custom 10-camera rig in 120FPS. The dataset contains 96 high-quality scenes showing various visual effects and human interactions in outdoor scenes. We develop a new algorithm, Deep 3D Mask Volume, which enables temporally-stable view extrapolation from binocular videos of dynamic scenes, captured by static cameras. Our algorithm addresses the temporal inconsistency of disocclusions by identifying the error-prone areas with a 3D mask volume, and replaces them with static background observed throughout the video. Our method enables manipulation in 3D space as opposed to simple 2D masks. We demonstrate better temporal stability than frame-by-frame static view synthesis methods, or those that use 2D masks. The resulting view synthesis videos show minimal flickering artifacts and allow for larger translational movements.

We present a novel solution to the garment animation problem through deep learning. Our contribution allows animating any template outfit with arbitrary topology and geometric complexity. Recent works develop models for garment edition, resizing and animation at the same time by leveraging the support body model (encoding garments as body homotopies). This leads to complex engineering solutions that suffer from scalability, applicability and compatibility. By limiting our scope to garment animation only, we are able to propose a simple model that can animate any outfit, independently of its topology, vertex order or connectivity. Our proposed architecture maps outfits to animated 3D models into the standard format for 3D animation (blend weights and blend shapes matrices), automatically providing of compatibility with any graphics engine. We also propose a methodology to complement supervised learning with an unsupervised physically based learning that implicitly solves collisions and enhances cloth quality.

The scale of deep learning nowadays calls for efficient distributed training algorithms. Decentralized momentum SGD (DmSGD), in which each node averages only with its neighbors, is more communication efficient than vanilla Parallel momentum SGD that incurs global average across all computing nodes. On the other hand, the large-batch training has been demonstrated critical to achieve runtime speedup. This motivates us to investigate how DmSGD performs in the large-batch scenario. In this work, we find the momentum term can amplify the inconsistency bias in DmSGD. Such bias becomes more evident as batch-size grows large and hence results in severe performance degradation. We next propose DecentLaM, a novel decentralized large-batch momentum SGD to remove the momentum-incurred bias. The convergence rate for both strongly convex and non-convex scenarios is established. Our theoretical results justify the superiority of DecentLaM to DmSGD especially in the large-batch scenario. Experimental results on a a variety of computer vision tasks and models show that DecentLaM promises both efficient and high-quality training.