ICCV 2021 Papers

The performance of surface registration relies heavily on the metric used for the alignment error between the source and target shapes. Traditionally, such a metric is based on the point-to-point or point-to-plane distance from the points on the source surface to their closest points on the target surface, which is susceptible to failure due to instability of the closest-point correspondence. In this paper, we propose a novel metric based on the intersection points between the two shapes and a random straight line, which does not assume a specific correspondence. We verify the effectiveness of this metric by extensive experiments, including its direct optimization for a single registration problem as well as unsupervised learning for a set of registration problems. The results demonstrate that the algorithms utilizing our proposed metric outperforms the state-of-the-art optimization-based and unsupervised learning-based methods.

Nowadays modern displays are capable to render video content with high dynamic range (HDR) and wide color gamut (WCG). However, most available resources are still in standard dynamic range (SDR). Therefore, there is an urgent demand to transform existing SDR-TV contents into their HDR-TV versions. In this paper, we conduct an analysis of SDRTV-to-HDRTV task by modeling the formation of SDRTV/HDRTV content. Base on the analysis, we propose a three-step solution pipeline including adaptive global color mapping, local enhancement and highlight generation. Moreover, the above analysis inspires us to present a lightweight network that utilizes global statistics as guidance to conduct image-adaptive color mapping. In addition, we construct a dataset using HDR videos in HDR10 standard, named HDRTV1K, and select five metrics to evaluate the results of SDRTV-to-HDRTV algorithms. Furthermore, our final results achieve state-of-the-art performance in quantitative comparisons and visual quality. The code and dataset are available at https://github.com/chxy95/HDRTVNet.

Most existing few-shot classification methods only consider generalization on one dataset (i.e., single-domain), failing to transfer across various seen and unseen domains. In this paper, we consider the more realistic multi-domain few-shot classification problem to investigate the cross-domain generalization. Two challenges exist in this new setting: (1) how to efficiently generate multi-domain feature representation, and (2) how to explore domain correlations for better cross-domain generalization. We propose a parameter-efficient multi-mode modulator to address both challenges. First, the modulator is designed to maintain multiple modulation parameters (one for each domain) in a single network, thus achieving single-network multi-domain representation. Given a particular domain, domain-aware features can be efficiently generated with the well-devised separative selection module and cooperative query module. Second, we further divide the modulation parameters into the domain-specific set and the domain-cooperative set to explore the intra-domain information and inter-domain correlations, respectively. The intra-domain information describes each domain independently to prevent negative interference. The inter-domain correlations guide information sharing among relevant domains to enrich their own representation. Moreover, unseen domains can utilize the correlations to obtain an adaptive combination of seen domains for extrapolation. We demonstrate that the proposed multi-mode modulator achieves state-of-the-art results on the challenging META-DATASET benchmark, especially for unseen test domains.

We consider the scalable recognition problem in the fine-grained expert domain where large-scale data collection is easy whereas annotation is difficult. Existing solutions are typically based on semi-supervised or self-supervised learning. We propose an alternative new framework, MEMORABLE, based on machine teaching and online crowdsourcing platforms. A small amount of data is first labeled by experts and then used to teach online annotators for the classes of interest, who finally label the entire dataset. Preliminary studies show that the accuracy of classifiers trained on the final dataset is a function of the accuracy of the student annotators. A new machine teaching algorithm, CMaxGrad, is then proposed to enhance this accuracy by introducing explanations in a state-of-the-art machine teaching algorithm. For this, CMaxGrad leverages counterfactual explanations, which take into account student predictions, thereby proving feedback that is student-specific, explicitly addresses the causes of student confusion, and adapts to the level of competence of the student. Experiments show that both MEMORABLE and CMaxGrad outperform existing solutions to their respective problems.

Every time you sit in front of a TV or monitor, your face is actively illuminated by time-varying patterns of light. This paper proposes to use this time-varying illumination for synthetic relighting of your face with any new illumination condition. In doing so, we take inspiration from the light stage work of Debevec et al. [4], who first demonstrated the ability to relight people captured in a controlled lighting environment. Whereas existing light stages require expensive, room-scale spherical capture gantries and exist in only a few labs in the world, we demonstrate how to acquire useful data from a normal TV or desktop monitor. Instead of subjecting the user to uncomfortable rapidly flashing light patterns, we operate on images of the user watching a YouTube video or other standard content. We train a deep network on images plus monitor patterns of a given user and learn to predict images of that user under any target illumination (monitor pattern). Experimental evaluation shows that our method produces realistic relighting results.

Gradient-based meta-learning relates task-specific models to a meta-model by gradients. By this design, an algorithm first optimizes the task-specific models by an inner loop and then backpropagates meta-gradients through the loop to update the meta-model. The number of inner-loop optimization steps has to be small (e.g., one step) to avoid high-order derivatives, big memory footprints, and the risk of vanishing or exploding meta-gradients. We propose an intuitive teacher-student scheme to enable the gradient-based meta-learning algorithms to explore long horizons by the inner loop. The key idea is to employ a student network to adequately explore the search space of task-specific models (e.g., by more than ten steps), and a teacher then takes a "leap" toward the regions probed by the student. The teacher not only arrives at a high-quality model but also defines a lightweight computation graph for meta-gradients. Our approach is generic; it performs well when applied to four meta-learning algorithms over three tasks: few-shot learning, long-tailed classification, and meta-attack.

High quality facial image editing is a challenging problem in the movie post-production industry, requiring a high degree of control and identity preservation. Previous works that attempt to tackle this problem may suffer from the entanglement of facial attributes and the loss of the person's identity. Furthermore, many algorithms are limited to a certain task. To tackle these limitations, we propose to edit facial attributes via the latent space of a StyleGAN generator, by training a dedicated latent transformation network and incorporating explicit disentanglement and identity preservation terms in the loss function. We further introduce a pipeline to generalize our face editing to videos. Our model achieves a disentangled, controllable, and identity-preserving facial attribute editing, even in the challenging case of real (i.e., non-synthetic) images and videos. We conduct extensive experiments on image and video datasets and show that our model outperforms other state-of-the-art methods in visual quality and quantitative evaluation. Source codes are available at https://github.com/InterDigitalInc/latent-transformer.

In this paper, we propose HF2-VAD, a Hybrid framework that integrates Flow reconstruction and Frame prediction seamlessly to handle Video Anomaly Detection. Firstly, we design the network of ML-MemAE-SC (Multi-Level Memory modules in an Autoencoder with Skip Connections) to memorize normal patterns for optical flow reconstruction so that abnormal events can be sensitively identified with larger flow reconstruction errors. More importantly, conditioned on the reconstructed flows, we then employ a Conditional Variational Autoencoder (CVAE), which captures the high correlation between video frame and optical flow, to predict the next frame given several previous frames. By CVAE, the quality of flow reconstruction essentially influences that of frame prediction. Therefore, poorly reconstructed optical flows of abnormal events further deteriorate the quality of the final predicted future frame, making the anomalies more detectable. Experimental results demonstrate the effectiveness of the proposed method. Code is available at https://github.com/LiUzHiAn/hf2vad.

Scanning transmission electron microscopy (STEM) is a powerful technique in high-resolution atomic imaging of materials. Decreasing scanning time and reducing electron beam exposure with an acceptable signal-to-noise results are two popular research aspects when applying STEM to beam-sensitive materials. Specifically, partially sampling with fixed electron doses is one of the most important solutions, and then the lost information is restored by computational methods. Following successful applications of deep learning in image in-painting, we have developed an encoder-decoder network to reconstruct STEM images in extremely sparse sampling case. In our model, we combine both local pixel information from convolution operators and global texture features, by applying specific filter operations on frequency domain to acquire initial reconstruction and global structure prior. Our method can effectively restore texture structures and be robust in different sampling ratios with Poisson noise. A comprehensive study demonstrates that our method gains about 50% performance enhancement in comparison with the state-of-art methods. Code is available at https://github.com/icthrm/Sparse-Sampling-Reconstruction.

Predicting the future frames of a video is a challenging task, in part due to the underlying stochastic real-world phenomena. Prior approaches to solve this task typically estimate a latent prior characterizing this stochasticity, however do not account for the predictive uncertainty of the (deep learning) model. Such approaches often derive the training signal from the mean-squared error (MSE) between the generated frame and the ground truth, which can lead to sub-optimal training, especially when the predictive uncertainty is high. Towards this end, we introduce Neural Uncertainty Quantifier (NUQ) - a stochastic quantification of the model's predictive uncertainty, and use it to weigh the MSE loss. We propose a hierarchical, variational framework to derive NUQ in a principled manner using a deep, Bayesian graphical model. Our experiments on three benchmark stochastic video prediction datasets show that our proposed framework trains more effectively compared to the state-of-the-art models (especially when the training sets are small), while demonstrating better video generation quality and diversity against several evaluation metrics.

Anomaly detection, the task of identifying unusual samples in data, often relies on a large set of training samples. In this work, we consider the setting of few-shot anomaly detection in images, where only a few images are given at training. We devise a hierarchical generative model that captures the multi-scale patch distribution of each training image. We further enhance the representation of our model by using image transformations and optimize scale-specific patch-discriminators to distinguish between real and fake patches of the image, as well as between different transformations applied to those patches. The anomaly score is obtained by aggregating the patch-based votes of the correct transformation across scales and image regions. We demonstrate the superiority of our method on both the one-shot and few-shot settings, on the datasets of Paris, CIFAR10, MNIST and FashionMNIST as well as in the setting of defect detection on MVTec. In all cases, our method outperforms the recent baseline methods.

Recent progress in multi-object tracking (MOT) has shown great significance of a robust scoring mechanism for potential tracks. However, the lack of available data in MOT makes it difficult to learn a general scoring mechanism. Multiple cues including appearance, motion and etc., are limitedly utilized in current manual scoring functions. In this paper, we propose a Multiple Nodes Tracking (MNT) framework that adapts to most trackers. Based on this framework, a Recurrent Tracking Unit (RTU) is designed to score potential tracks through long-term information. In addition, we present a method of generating simulated tracking data without real data to overcome the defect of limited available data in MOT. The experiments demonstrate that our simulated tracking data is effective for training RTU and achieves state-of-the-art performance on both MOT17 and MOT16 benchmarks. Meanwhile, RTU can be flexibly plugged into classic trackers such as DeepSORT and MHT, and makes remarkable improvements as well.

Casual photography is often performed in uncontrolled lighting that can result in low quality images and degrade the performance of downstream processing. We consider the problem of estimating surface normal and reflectance maps of scenes depicting people despite these conditions by supplementing the available visible illumination with a single near infrared (NIR) light source and camera, a so-called "dark flash image". Our method takes as input a single color image captured under arbitrary visible lighting and a single dark flash image captured under controlled front-lit NIR lighting at the same viewpoint, and computes a normal map, a diffuse albedo map, and a specular intensity map of the scene. Since ground truth normal and reflectance maps of faces are difficult to capture, we propose a novel training technique that combines information from two readily available and complementary sources: a stereo depth signal and photometric shading cues. We evaluate our method over a range of subjects and lighting conditions and describe two applications: optimizing stereo geometry and filling the shadows in an image.

In this paper, we introduce a deep multi-view stereo (MVS) system that jointly predicts depths, surface normals and per-view confidence maps. The key to our approach is a novel solver that iteratively solves for per-view depth map and normal map by optimizing an energy potential based upon the local planar assumption. Specifically, the algorithm updates depth map by propagating from neighboring pixels with slanted planes, and updates normal map with local probabilistic plane fitting. Both two steps are monitored by a customized confidence map. This confidence-based solver is not only effective as a post-processing tool for plane based depth refinement and completion, but also differentiable such that it can be efficiently integrated into deep learning pipelines. Our multi-view stereo system employs multiple optimization steps of the solver over the initial prediction of depths and surface normals. The whole system can be trained end-to-end, decoupling the challenging problem of matching pixels within poorly textured regions from the cost volume based neural network. Experimental results on ScanNet and RGB-D Scenes V2 demonstrate state-of-the-art performance of the proposed deep MVS system on multi-view depth estimation, with our proposed solver consistently improving the depth quality over both conventional and deep learning based MVS pipelines.

We consider the deep point cloud analysis tasks where the inputs of the networks are randomly rotated. Recent progress in rotation-invariant point cloud analysis is mainly driven by converting point clouds into their respective canonical poses, and principal component analysis (PCA) is a practical tool to achieve this. Due to the imperfect alignment of PCA, most of the current works are devoted to developing powerful network structures and features to overcome this deficiency, without thoroughly analyzing the PCA-based canonical poses themselves. In this work, we present a detailed study w.r.t. the PCA-based canonical poses of point clouds. Our investigation reveals that the ambiguity problem associated with the PCA-based canonical poses is handled insufficiently in some recent works. To this end, we develop a simple pose selector module for disambiguation, which presents noticeable enhancement (i.e., 5:3% classification accuracy) over state-of-the-art approaches on the challenging real-world dataset.

Tremendous progress has been made in visual representation learning, notably with the recent success of self-supervised contrastive learning methods. Supervised contrastive learning has also been shown to outperform its cross-entropy counterparts by leveraging labels for choosing where to contrast. However, there has been little work to explore the transfer capability of contrastive learning to a different domain. In this paper, we conduct a comprehensive study on the transferability of learned representations of different contrastive approaches for linear evaluation, full-network transfer, and few-shot recognition on 12 downstream datasets from different domains, and object detection tasks on MSCOCO and VOC0712. The results show that the contrastive approaches learn representations that are easily transferable to a different downstream task. We further observe that the joint objective of self-supervised contrastive loss with cross-entropy/supervised-contrastive loss leads to better transferability of these models over their supervised counterparts. Our analysis reveals that the representations learned from the contrastive approaches contain more low/mid-level semantics than cross-entropy models, which enables them to quickly adapt to a new task. Our codes and models will be publicly available to facilitate future research on transferability of visual representations.

Vulnerability of 3D point cloud (PC) classifiers has become a grave concern due to the popularity of 3D sensors in safety-critical applications. Existing adversarial attacks against 3D PC classifiers are all test-time evasion (TTE) attacks that aim to induce test-time misclassifications using knowledge of the classifier. But since the victim classifier is usually not accessible to the attacker, the threat is largely diminished in practice, as PC TTEs typically have poor transferability. Here, we propose the first backdoor attack (BA) against PC classifiers. Originally proposed for images, BAs poison the victim classifier's training set so that the classifier learns to decide to the attacker's target class whenever the attacker's backdoor pattern is present in a given input sample. Significantly, BAs do not require knowledge of the victim classifier. Different from image BAs, we propose to insert a cluster of points into a PC as a robust backdoor pattern customized for 3D PCs. Such clusters are also consistent with a physical attack (i.e., with a captured object in a scene). We optimize the cluster's location using an independently trained surrogate classifier and choose the cluster's local geometry to evade possible PC preprocessing and PC anomaly detectors (ADs). Experimentally, our BA achieves a uniformly high success rate (>=87%) and shows evasiveness against state-of-the-art PC ADs. Code is available at https://github.com/zhenxianglance/PCBA.

Tracking non-rigidly deforming scenes using range sensors has numerous applications including computer vision, AR/VR, and robotics. However, due to occlusions and physical limitations of range sensors, existing methods only handle the visible surface, thus causing discontinuities and incompleteness in the motion field. To this end, we introduce 4DComplete, a novel data-driven approach that estimates the non-rigid motion for the unobserved geometry. 4DComplete takes as input a partial shape and motion observation, extracts 4D time-space embedding, and jointly infers the missing geometry and motion field using a sparse fully-convolutional network. For network training, we constructed a large-scale synthetic dataset called DeformingThings4D, which consists of 1,972 animation sequences spanning 31 different animals or humanoid categories with dense 4D annotation. Experiments show that 4DComplete 1) reconstructs high-resolution volumetric shape and motion field from a partial observation, 2) learns an entangled 4D feature representation that benefits both shape and motion estimation, 3) yields more accurate and natural deformation than classic non-rigid priors such as As-RigidAs-Possible (ARAP) deformation, and 4) generalizes well to unseen objects in real-world sequences.

We present 4D-Net, a 3D object detection approach, which utilizes 3D Point Cloud and RGB sensing information, both in time. We are able to incorporate the 4D information by performing a novel dynamic connection learning across various feature representations and levels of abstraction and by observing geometric constraints. Our approach outperforms the state-of-the-art and strong baselines on the Waymo Open Dataset. 4D-Net is better able to use motion cues and dense image information to detect distant objects more successfully. We will open source the code.