ICCV 2021 Papers

In this paper, we present a new tracking architecture with an encoder-decoder transformer as the key component. The encoder models the global spatio-temporal feature dependencies between target objects and search regions, while the decoder learns a query embedding to predict the spatial positions of the target objects. Our method casts object tracking as a direct bounding box prediction problem, without using any proposals or predefined anchors. With the encoder-decoder transformer, the prediction of objects just uses a simple fully-convolutional network, which estimates the corners of objects directly. The whole method is end-to-end, does not need any postprocessing steps such as cosine window and bounding box smoothing, thus largely simplifying existing tracking pipelines. The proposed tracker achieves state-of-the-art performance on multiple challenging short-term and long-term benchmarks, while running at real-time speed, being 6x faster than Siam R-CNN. Code and models are open-sourced at https://github.com/researchmm/Stark.

Various deep learning techniques have been proposed to solve the single-view 2D-to-3D pose estimation problem. While the average prediction accuracy has been improved significantly over the years, the performance on hard poses with depth ambiguity, self-occlusion, and complex or rare poses is still far from satisfactory. In this work, we target these hard poses and present a novel skeletal GNN learning solution. To be specific, we propose a hop-aware hierarchical channel-squeezing fusion layer to effectively extract relevant information from neighboring nodes while suppressing undesired noises in GNN learning. In addition, we propose a temporal-aware dynamic graph construction procedure that is robust and effective for 3D pose estimation. Experimental results on the Human3.6M dataset show that our solution achieves a 10.3% average prediction accuracy improvement and greatly improves on hard poses over state-of-the-art techniques. We further apply the proposed technique on the skeleton-based action recognition task and also achieve state-of-the-art performance.

Recent works on implicit neural representations have shown promising results for multi-view surface reconstruction. However, most approaches are limited to relatively simple geometries and usually require clean object masks for reconstructing complex and concave objects. In this work, we introduce a novel neural surface reconstruction framework that leverages the knowledge of stereo matching and feature consistency to optimize the implicit surface representation. More specifically, we apply a signed distance field (SDF) and a surface light field to represent the scene geometry and appearance respectively. The SDF is directly supervised by geometry from stereo matching, and is refined by optimizing the multi-view feature consistency and the fidelity of rendered images. Our method is able to improve the robustness of geometry estimation and support reconstruction of complex scene topologies. Extensive experiments have been conducted on DTU, EPFL and Tanks and Temples datasets. Compared to previous state-of-the-art methods, our method achieves better mesh reconstruction in wide open scenes without masks as input.

Spatio-temporal convolution often fails to learn motion dynamics in videos and thus an effective motion representation is required for video understanding in the wild. In this paper, we propose a rich and robust motion representation based on spatio-temporal self-similarity (STSS). Given a sequence of frames, STSS represents each local region as similarities to its neighbors in space and time. By converting appearance features into relational values, it enables the learner to better recognize structural patterns in space and time. We leverage the whole volume of STSS and let our model learn to extract an effective motion representation from it. The proposed neural block, dubbed SELFY, can be easily inserted into neural architectures and trained end-to-end without additional supervision. With a sufficient volume of the neighborhood in space and time, it effectively captures long-term interaction and fast motion in the video, leading to robust action recognition. Our experimental analysis demonstrates its superiority over previous methods for motion modeling as well as its complementarity to spatio-temporal features from direct convolution. On the standard action recognition benchmarks, SomethingSomething-V1 & V2, Diving-48, and FineGym, the proposed method achieves the state-of-the-art results.

We propose a new method to detect deepfake images using the cue of the source feature inconsistency within the forged images. It is based on the hypothesis that images' distinct source features can be preserved and extracted after going through state-of-the-art deepfake generation processes. We introduce a novel representation learning approach, called pair-wise self-consistency learning (PCL), for training ConvNets to extract these source features and detect deepfake images. It is accompanied by a new image synthesis approach, called inconsistency image generator (I2G), to provide richly annotated training data for PCL. Experimental results on seven popular datasets show that our models improve averaged AUC from 96.45% to 98.05% over the state of the art in the in-dataset evaluation and from 86.03% to 92.18% in the cross-dataset evaluation.

Synthesizing images of a person in novel poses from a single image is a highly ambiguous task. Most existing approaches require paired training images; i.e. images of the same person with the same clothing in different poses. However, obtaining sufficiently large datasets with paired data is challenging and costly. Previous methods that forego paired supervision lack realism. We propose a self-supervised framework named SPICE (Self-supervised Person Image CrEation) that closes the image quality gap with supervised methods. The key insight enabling self-supervision is to exploit 3D information about the human body in several ways. First, the 3D body shape must remain unchanged when reposing. Second, representing body pose in 3D enables reasoning about self occlusions. Third, 3D body parts that are visible before and after reposing, should have similar appearance features. Once trained, SPICE takes an image of a person and generates a new image of that person in a new target pose. SPICE achieves state-of-the-art performance on the DeepFashion dataset, improving the FID score from 29.9 to 7.8 compared with previous unsupervised methods, and with performance similar to the state-of-the-art supervised method (6.4). SPICE also generates temporally coherent videos given an input image and a sequence of poses, despite being trained on static images only.

Many real-world ML deployments face the challenge of training a rare category model with a small labeling bud- get. In these settings, there is often access to large amounts of unlabeled data, therefore it is attractive to consider semi-supervised or active learning approaches to reduce human labeling effort. However, prior approaches make two assumptions that do not often hold in practice; (a) one has access to a modest amount of labeled data to bootstrap learning and (b) every image belongs to a common category of interest. In this paper, we consider the scenario where we start with as-little-as five labeled positives of a rare category and a large amount of unlabeled data of which 99.9% of it is negatives. We propose an active semi-supervised method for building accurate models in this challenging setting. Our method leverages two key ideas: (a) Utilize human and machine effort where they are most effective; human labels are used to identify "needle-in-a-haystack" positives, while machine-generated pseudo-labels are used to identify negatives. (b) Adapt recently proposed representation learning techniques for handling extremely imbalanced human labeled data to iteratively train models with noisy machine labeled data. We compare our approach with prior active learning and semi-supervised approaches, demonstrating significant improvements in accuracy per unit labeling effort, particularly on a tight labeling budget.

Learning RAW-to-sRGB mapping has drawn increasing attention in recent years, wherein an input raw image is trained to imitate the target sRGB image captured by another camera. However, the severe color inconsistency makes it very challenging to generate well-aligned training pairs of input raw and target sRGB images. While learning with inaccurately aligned supervision is prone to causing pixel shift and producing blurry results. In this paper, we circumvent such issue by presenting a joint learning model for image alignment and RAW-to-sRGB mapping. To diminish the effect of color inconsistency in image alignment, we introduce to use a global color mapping (GCM) module to generate an initial sRGB image given the input raw image, which can keep the spatial location of the pixels unchanged, and the target sRGB image is utilized to guide GCM for converting the color towards it. Then a pre-trained optical flow estimation network (e.g., PWC-Net) is deployed to warp the target sRGB image to align with the GCM output. To alleviate the effect of inaccurately aligned supervision, the warped target sRGB image is leveraged to learn RAW-to-sRGB mapping. When training is done, the GCM module and optical flow network can be detached, thereby bringing no extra computation cost for inference. Experiments show that our method performs favorably against state-of-the-arts on ZRR and SR-RAW datasets. With our joint learning model, a light-weight backbone can achieve better quantitative and qualitative performance on ZRR dataset. Codes are available at https://github.com/cszhilu1998/RAW-to-sRGB.

The widespread use of always-connected digital cameras in our everyday life has led to increasing concerns about the users' privacy and security. How to develop privacy-preserving computer vision systems? In particular, we want to prevent the camera from obtaining detailed visual data that may contain private information. However, we also want the camera to capture useful information to perform computer vision tasks. Inspired by the trend of jointly designing optics and algorithms, we tackle the problem of privacy-preserving human pose estimation by optimizing an optical encoder (hardware-level protection) with a software decoder (convolutional neural network) in an end-to-end framework. We introduce a visual privacy protection layer in our optical encoder that, parametrized appropriately, enables the optimization of the camera lens's point spread function (PSF). We validate our approach with extensive simulations and a prototype camera. We show that our privacy-preserving deep optics approach successfully degrades or inhibits private attributes while maintaining important features to perform human pose estimation.

Implicit neural rendering techniques have shown promising results for novel view synthesis. However, existing methods usually encode the entire scene as a whole, which is generally not aware of the object identity and limits the ability to the high-level editing tasks such as moving or adding furniture. In this paper, we present a novel neural scene rendering system, which learns an object-compositional neural radiance field and produces realistic rendering with editing capability for a clustered and real-world scene. Specifically, we design a novel two-pathway architecture, in which the scene branch encodes the scene geometry and appearance, and the object branch encodes each standalone object conditioned on learnable object activation codes. To survive the training in heavily cluttered scenes, we propose a scene-guided training strategy to solve the 3D space ambiguity in the occluded regions and learn sharp boundaries for each object. Extensive experiments demonstrate that our system not only achieves competitive performance for static scene novel-view synthesis, but also produces realistic rendering for object-level editing.

We propose a novel loss weighting algorithm, called loss scale balancing (LSB), for multi-task learning (MTL) of pixelwise vision tasks. An MTL model is trained to estimate multiple pixelwise predictions using an overall loss, which is a linear combination of individual task losses. The proposed algorithm dynamically adjusts the linear weights to learn all tasks effectively. Instead of controlling the trend of each loss value directly, we balance the loss scale --- the product of the loss value and its weight --- periodically. In addition, by evaluating the difficulty of each task based on the previous loss record, the proposed algorithm focuses more on difficult tasks during training. Experimental results show that the proposed algorithm outperforms conventional weighting algorithms for MTL of various pixelwise tasks. Codes are available at https://github.com/jaehanlee-mcl/LSB-MTL.

Absolute camera pose regression methods estimate the position and orientation of a camera by only using the captured image. A convolutional backbone with a multi-layer perceptron head is trained with images and pose labels to embed a single reference scene at a time. Recently, this framework was extended for learning multiple scenes with a single model by adding a multi-layer perceptron head per scene. In this work, we propose to learn multi-scene absolute camera pose regression with transformers, where encoders are used to aggregate activation maps with self-attention and deocoders transform latent features into candidate pose predictions in parallel, each associated with a different scene. This formulation allows our model to focus on general features that are informative for localization while embedding multiple scenes at once. We evaluate our method on commonly benchmarked indoor and outdoor datasets and show that it surpasses both multi-scene and single-scene absolute pose regressors.

How to make the appearance and motion information interact effectively to accommodate complex scenarios is a fundamental issue in flow-based zero-shot video object segmentation. In this paper, we propose an Attentive Multi-Modality Collaboration Network (AMC-Net) to utilize appearance and motion information uniformly. Specifically, AMC-Net fuses robust information from multi-modality features and promotes their collaboration in two stages. First, we propose a Multi-Modality Co-Attention Gate (MCG) on the bilateral encoder branches, in which a gate function is used to formulate co-attention scores for balancing the contributions of multi-modality features and suppressing the redundant and misleading information. Then, we propose a Motion Correction Module (MCM) with a visual-motion attention mechanism, which is constructed to emphasize the features of foreground objects by incorporating the spatio-temporal correspondence between appearance and motion cues. Extensive experiments on three public challenging benchmark datasets verify that our proposed network performs favorably against existing state-of-the-art methods via training with fewer data.

Recovering high-quality 3D human motion in complex scenes from monocular videos is important for many applications, ranging from AR/VR to robotics. However, capturing realistic human-scene interactions, while dealing with occlusions and partial views, is challenging; current approaches are still far from achieving compelling results. We address this problem by proposing LEMO: LEarning human MOtion priors for 4D human body capture. By leveraging the large-scale motion capture dataset AMASS, we introduce a novel motion smoothness prior, which strongly reduces the jitters exhibited by poses recovered over a sequence. Furthermore, to handle contacts and occlusions occurring frequently in body-scene interactions, we design a contact friction term and a contact-aware motion infiller obtained via per-instance self-supervised training. To prove the effectiveness of the proposed motion priors, we combine them into a novel pipeline for 4D human body capture in 3D scenes. With our pipeline, we demonstrate high-quality 4D human body capture, reconstructing smooth motions and physically plausible body-scene interactions. The code and data are available at https://sanweiliti.github.io/LEMO/LEMO.html.

Currently, the state-of-the-art methods treat few-shot semantic segmentation task as a conditional foreground-background segmentation problem, assuming each class is independent. In this paper, we introduce the concept of meta-class, which is the meta information (e.g. certain middle-level features) shareable among all classes. To explicitly learn meta-class representations in few-shot segmentation task, we propose a novel Meta-class Memory based few-shot segmentation method (MM-Net), where we introduce a set of learnable memory embeddings to memorize the meta-class information during the base class training and transfer to novel classes during the inference stage. Moreover, for the k-shot scenario, we propose a novel image quality measurement module to select images from the set of support images. A high-quality class prototype could be obtained with the weighted sum of support image features based on the quality measure. Experiments on both PASCAL-5^i and COCO datasets show that our proposed method is able to achieve state-of-the-art results in both 1-shot and 5-shot settings. Particularly, our proposed MM-Net achieves 37.5% mIoU on the COCO dataset in 1-shot setting, which is 5.1% higher than the previous state-of-the-art.

Existing differentiable neural architecture search approaches simply assume the architectural distribution on each edge is independent of each other, which conflicts with the intrinsic properties of architecture. In this paper, we view the architectural distribution as the latent representation of specific data points. Then we propose Variational Information Maximization Neural Architecture Search (VIM-NAS) to leverage a simple but effective convolutional neural network to model the latent representation, and optimize for a tractable variational lower bound to the mutual information between the data points and the latent representations. VIM-NAS automatically learns a near one-hot distribution from a continuous distribution with extremely fast convergence speed, e.g., converging with one epoch. Experimental results demonstrate VIM-NAS achieves state-of-the-art performance on various search spaces, including DARTS search space, NAS-Bench-1shot1, NAS-Bench-201, and simplified search spaces S1-S4. Specifically, VIM-NAS achieves a top-1 error rate of 2.45% and 15.80% within 10 minutes on CIFAR-10 and CIFAR-100, respectively, and a top-1 error rate of 24.0% when transferred to ImageNet.

Person re-identification (Re-ID) aims to match pedestrians under dis-joint cameras. Most Re-ID methods formulate it as visual representation learning and image search, and its accuracy is consequently affected greatly by the search space. Spatial-temporal information has been proven to be efficient to filter irrelevant negative samples and significantly improve Re-ID accuracy. However, existing spatial-temporal person Re-ID methods are still rough and do not exploit spatial-temporal information sufficiently. In this paper, we propose a novel instance-level and spatial-temporal disentangled Re-ID method (InSTD), to improve Re-ID accuracy. In our proposed framework, personalized information such as moving direction is explicitly considered to further narrow down the search space. Besides, the spatial-temporal transferring probability is disentangled from joint distribution to marginal distribution, so that outliers can also be well modeled. Abundant experimental analyses on two datasets are presented, which demonstrates the superiority and provides more insights into our method. The proposed method achieves mAP of 90.8% on Market-1501 and 89.1% on DukeMTMC-reID, improving from the baseline 82.2% and 72.7%, respectively. Besides, in order to provide a better benchmark for person re-identification, we release a cleaned data list of DukeMTMC-reID with this paper: https://github.com/RenMin1991/cleaned-DukeMTMC-reID.

The prevalence of relation networks in computer vision is in stark contrast to underexplored point-based methods. In this paper, we explore the possibilities of local relation operators and survey their feasibility. We propose a scalable and efficient module, called group relation aggregator. The module computes a feature of a group based on the aggregation of the features of the inner-group points weighted by geometric relations and semantic relations. For convenience, we generalize groupwise operations to assemble this module. We adopt this module to design our RPNet. We further verify the expandability of RPNet, in terms of both depth and width, on the tasks of classification and segmentation. Surprisingly, empirical results show that wider RPNet fits for classification, while deeper RPNet works better on segmentation. RPNet achieves state-of-the-art for classification and segmentation on challenging benchmarks. We also compare our local aggregator with PointNet++, with around 30% parameters and 50% computation saving. Finally, we conduct experiments to reveal the robustness of RPNet with regard to rigid transformation and noises.

In this work, we address the problem of jointly estimating albedo, normals, depth and 3D spatially-varying lighting from a single image. Most existing methods formulate the task as image-to-image translation, ignoring the 3D properties of the scene. However, indoor scenes contain complex 3D light transport where a 2D representation is insufficient. In this paper, we propose a unified, learning-based inverse rendering framework that formulates 3D spatially-varying lighting. Inspired by classic volume rendering techniques, we propose a novel Volumetric Spherical Gaussian representation for lighting, which parameterizes the exitant radiance of the 3D scene surfaces on a voxel grid. We design a physicsbased differentiable renderer that utilizes our 3D lighting representation, and formulates the energy-conserving image formation process that enables joint training of all intrinsic properties with the re-rendering constraint. Our model ensures physically correct predictions and avoids the need for ground-truth HDR lighting which is not easily accessible. Experiments show that our method outperforms prior works both quantitatively and qualitatively, and is capable of producing photorealistic results for AR applications such as virtual object insertion even for highly specular objects.