CVPR 2022 Papers

User-level differential privacy (DP) provides certifiable privacy guarantees to the information that is specific to any user's data in federated learning. Existing methods that ensure user-level DP come at the cost of severe accuracy decrease. In this paper, we study the cause of model performance degradation in federated learning with user-level DP guarantee. We find the key to solving this issue is to naturally restrict the norm of local updates before executing operations that guarantee DP. To this end, we propose two techniques, Bounded Local Update Regularization and Local Update Sparsification, to increase model quality without sacrificing privacy. We provide theoretical analysis on the convergence of our framework and give rigorous privacy guarantees. Extensive experiments show that our framework significantly improves the privacy-utility trade-off over the state-of-the-arts for federated learning with user-level DP guarantee.

We propose Differentiable Stereopsis, a multi-view stereo approach that reconstructs shape and texture from few input views and noisy cameras. We pair traditional stereopsis and modern differentiable rendering to build an end-to-end model which predicts textured 3D meshes of objects with varying topologies and shape. We frame stereopsis as an optimization problem and simultaneously update shape and cameras via simple gradient descent. We run an extensive quantitative analysis and compare to traditional multi-view stereo techniques and state-of-the-art learning based methods. We show compelling reconstructions on challenging real-world scenes and for an abundance of object types with complex shape, topology and texture.

We introduce DiffPhy, a differentiable physics-based model for articulated 3d human motion reconstruction from video. Applications of physics-based reasoning in human motion analysis have so far been limited, both by the complexity of constructing adequate physical models of articulated human motion, and by the formidable challenges of performing stable and efficient inference with physics in the loop. We jointly address such modeling and inference challenges by proposing an approach that combines a physically plausible body representation with anatomical joint limits, a differentiable physics simulator, and optimization techniques that ensure good performance and robustness to suboptimal local optima. In contrast to several recent methods, our approach readily supports full-body contact including interactions with objects in the scene. Most importantly, our model connects end-to-end with images, thus supporting direct gradient-based physics optimization by means of image-based loss functions. We validate the model by demonstrating that it can accurately reconstruct physically plausible 3d human motion from monocular video, both on public benchmarks with available 3d ground-truth, and on videos from the internet.

Current deep neural network approaches for camera pose estimation rely on scene structure for 3D motion estimation, but this decreases the robustness and thereby makes cross-dataset generalization difficult. In contrast, classical approaches to structure from motion estimate 3D motion utilizing optical flow and then compute depth. Their accuracy, however, depends strongly on the quality of the optical flow. To avoid this issue, direct methods have been proposed, which separate 3D motion from depth estimation but compute 3D motion using only image gradients in the form of normal flow. In this paper, we introduce a network NFlowNet, for normal flow estimation which is used to enforce robust and direct constraints. In particular, normal flow is used to estimate relative camera pose based on the cheirality (depth positivity) constraint. We achieve this by formulating the optimization problem as a differentiable cheirality layer, which allows for end-to-end learning of camera pose. We perform extensive qualitative and quantitative evaluation of the proposed DiffPoseNet's sensitivity to noise and its generalization across datasets. We compare our approach to existing state-of-the-art methods on KITTI, TartanAir, and TUM-RGBD datasets

Despite the popularity of Model Compression and Multitask Learning, how to effectively compress a multitask model has been less thoroughly analyzed due to the challenging entanglement of tasks in the parameter space. In this paper, we propose DiSparse, a simple, effective, and first-of-its-kind multitask pruning and sparse training scheme. We consider each task independently by disentangling the importance measurement and take the unanimous decisions among all tasks when performing parameter pruning and selection. Our experimental results demonstrate superior performance on various configurations and settings compared to popular sparse training and pruning methods. Besides the effectiveness in compression, DiSparse also provides a powerful tool to the multitask learning community. Surprisingly, we even observed better performance than some dedicated multitask learning methods in several cases despite the high model sparsity enforced by DiSparse. We analyzed the pruning masks generated with DiSparse and observed strikingly similar sparse network architecture identified by each task even before the training starts. We also observe the existence of a "watershed" layer where the task relatedness sharply drops, implying no benefits in continued parameters sharing. Our code and models will be available at: https://github.com/SHI-Labs/DiSparse-Multitask-Model-Compression.

Discriminative learning, restorative learning, and adversarial learning have proven beneficial for self-supervised learning schemes in computer vision and medical imaging. Existing efforts, however, omit their synergistic effects on each other in a ternary setup, which, we envision, can significantly benefit deep semantic representation learning. To realize this vision, we have developed DiRA, the first framework that unites discriminative, restorative, and adversarial learning in a unified manner to collaboratively glean complementary visual information from unlabeled medical images for fine-grained semantic representation learning. Our extensive experiments demonstrate that DiRA (1) encourages collaborative learning among three learning ingredients, resulting in more generalizable representation across organs, diseases, and modalities; (2) outperforms fully supervised ImageNet models and increases robustness in small data regimes, reducing annotation cost across multiple medical imaging applications; (3) learns fine-grained semantic representation, facilitating accurate lesion localization with only image-level annotation; and (4) enhances state-of-the-art restorative approaches, revealing that DiRA is a general mechanism for united representation learning. All code and pretrained models are available at https://github.com/JLiangLab/DiRA.

Evaluating photometric stereo using real-world dataset is important yet difficult. Existing datasets are insufficient due to their limited scale and random distributions in shape and material. This paper presents a new real-world photometric stereo dataset with "ground truth" normal maps, which is 10 times larger than the widely adopted one. More importantly, we propose to control the shape and material variations by fabricating objects from CAD models with carefully selected materials, covering typical aspects of reflectance properties that are distinctive for evaluating photometric stereo methods. By benchmarking recent photometric stereo methods using these 100 sets of images, with a special focus on recent learning based solutions, a 10 x 10 shape-material error distribution matrix is visualized to depict a "portrait" for each evaluated method. From such comprehensive analysis, open problems in this field are discussed. To inspire future research, this dataset is available at https://photometricstereo.github.io.

Shape implicit neural representations (INR) have recently shown to be effective in shape analysis and reconstruction tasks. Existing INRs require point coordinates to learn the implicit level sets of the shape. When a normal vector is available for each point, a higher fidelity representation can be learned, however normal vectors are often not provided as raw data. Furthermore, the method's initialization has been shown to play a crucial role for surface reconstruction. In this paper, we propose a divergence guided shape representation learning approach that does not require normal vectors as input. We show that incorporating a soft constraint on the divergence of the distance function favours smooth solutions that reliably orients gradients to match the unknown normal at each point, in some cases even better than approaches that use ground truth normal vectors directly. Additionally, we introduce a novel geometric initialization method for sinusoidal INRs that further improves convergence to the desired solution. We evaluate the effectiveness of our approach on the task of surface reconstruction and shape space learning and show SOTA performance compared to other unoriented methods.

Given a set of putative 3D-3D point correspondences, we aim to remove outliers and estimate rigid transformation with 6 degrees of freedom (DOF). Simultaneously estimating these 6 DOF is time-consuming due to high-dimensional parameter space. To solve this problem, it is common to decompose 6 DOF, i.e. independently compute 3-DOF rotation and 3-DOF translation. However, high non-linearity of 3-DOF rotation still limits the algorithm efficiency, especially when the number of correspondences is large. In contrast, we propose to decompose 6 DOF into (2+1) and (1+2) DOF. Specifically, (2+1) DOF represent 2-DOF rotation axis and 1-DOF displacement along this rotation axis. (1+2) DOF indicate 1-DOF rotation angle and 2-DOF displacement orthogonal to the above rotation axis. To compute these DOF, we design a novel two-stage strategy based on inlier set maximization. By leveraging branch and bound, we first search for (2+1) DOF, and then the remaining (1+2) DOF. Thanks to the proposed transformation decomposition and two-stage search strategy, our method is deterministic and leads to low computational complexity. We extensively compare our method with state-of-the-art approaches. Our method is more accurate and robust than the approaches that provide similar efficiency to ours. Our method is more efficient than the approaches whose accuracy and robustness are comparable to ours.

Group activity recognition is the task of understanding the activity conducted by a group of people as a whole in a multi-person video. Existing models for this task are often impractical in that they demand ground-truth bounding box labels of actors even in testing or rely on off-the-shelf object detectors. Motivated by this, we propose a novel model for group activity recognition that depends neither on bounding box labels nor on object detector. Our model based on Transformer localizes and encodes partial contexts of a group activity by leveraging the attention mechanism, and represents a video clip as a set of partial context embeddings. The embedding vectors are then aggregated to form a single group representation that reflects the entire context of an activity while capturing temporal evolution of each partial context. Our method achieves outstanding performance on two benchmarks, Volleyball and NBA datasets, surpassing not only the state of the art trained with the same level of supervision, but also some of existing models relying on stronger supervision.

In this paper, we present novel synthetic training data called self-blended images (SBIs) to detect deepfakes. SBIs are generated by blending pseudo source and target images from single pristine images, reproducing common forgery artifacts (e.g., blending boundaries and statistical inconsistencies between source and target images). The key idea behind SBIs is that more general and hardly recognizable fake samples encourage classifiers to learn generic and robust representations without overfitting to manipulation-specific artifacts. We compare our approach with state-of- the-art methods on FF++, CDF, DFD, DFDC, DFDCP, and FFIW datasets by following the standard cross-dataset and cross-manipulation protocols. Extensive experiments show that our method improves the model generalization to unknown manipulations and scenes. In particular, on DFDC and DFDCP where existing methods suffer from the domain gap between the training and test sets, our approach outperforms the baseline by 4.90% and 11.78% points in the cross-dataset evaluation, respectively. Code is available at https://github.com/mapooon/SelfBlendedImages.

Camouflaged object detection (COD) aims to identify objects that are perfectly embedded in their environment, which has various downstream applications in fields such as medicine, art, and agriculture. However, it is an extremely challenging task to spot camouflaged objects with the perception ability of human eyes. Hence, we claim that the goal of COD task is not just to mimic the human visual ability in a single RGB domain, but to go beyond the human biological vision. We then introduce the frequency domain as an additional clue to better detect camouflaged objects from backgrounds. To well involve the frequency clues into the CNN models, we present a powerful network with two special components. We first design a novel frequency enhancement module (FEM) to dig clues of camouflaged objects in the frequency domain. It contains the offline discrete cosine transform followed by the learnable enhancement. Then we use a feature alignment to fuse the features from RGB domain and frequency domain. Moreover, to further make full use of the frequency information, we propose the high-order relation module (HOR) to handle the rich fusion feature. Comprehensive experiments on three widely-used COD datasets show the proposed method significantly outperforms other state-of-the-art methods by a large margin. The code and results are released in https://github.com/luckybird1994/FDCOD.

Single image super-resolution (SISR) with generative adversarial networks (GAN) has recently attracted increasing attention due to its potentials to generate rich details. However, the training of GAN is unstable, and it often introduces many perceptually unpleasant artifacts along with the generated details. In this paper, we demonstrate that it is possible to train a GAN-based SISR model which can stably generate perceptually realistic details while inhibiting visual artifacts. Based on the observation that the local statistics (e.g., residual variance) of artifact areas are often different from the areas of perceptually friendly details, we develop a framework to discriminate between GAN-generated artifacts and realistic details, and consequently generate an artifact map to regularize and stabilize the model training process. Our proposed locally discriminative learning (LDL) method is simple yet effective, which can be easily plugged in off-the-shelf SISR methods and boost their performance. Experiments demonstrate that LDL outperforms the state-of-the-art GAN based SISR methods, achieving not only higher reconstruction accuracy but also superior perceptual quality on both synthetic and real-world datasets. Codes and models are available at https://github.com/csjliang/LDL.

A commonly observed failure mode of Neural Radiance Field (NeRF) is fitting incorrect geometries when given an insufficient number of input views. One potential reason is that standard volumetric rendering does not enforce the constraint that most of a scene's geometry consist of empty space and opaque surfaces. We formalize the above assumption through DS-NeRF (Depth-supervised Neural Radiance Fields), a loss for learning radiance fields that takes advantage of readily-available depth supervision. We leverage the fact that current NeRF pipelines require images with known camera poses that are typically estimated by running structure-from-motion (SFM). Crucially, SFM also produces sparse 3D points that can be used as "free" depth supervision during training: we add a loss to encourage the distribution of a ray's terminating depth matches a given 3D keypoint, incorporating depth uncertainty. DS-NeRF can render better images given fewer training views while training 2-3x faster. Further, we show that our loss is compatible with other recently proposed NeRF methods, demonstrating that depth is a cheap and easily digestible supervisory signal. And finally, we find that DS-NeRF can support other types of depth supervision such as scanned depth sensors and RGBD reconstruction outputs.

Existing approaches for Structure from Motion (SfM) produce impressive 3D reconstruction results especially when using imagery captured with large parallax. However, to create engaging video-content in movies and TV shows, the amount by which a camera can be moved while filming a particular shot is often limited. The resulting small-motion parallax between video frames makes standard geometry-based SfM approaches not as effective for movies and TV shows. To address this challenge, we propose a simple yet effective approach that uses single-frame depth-prior obtained from a pretrained network to significantly improve geometry-based SfM for our small-parallax setting. To this end, we first use the depth-estimates of the detected keypoints to reconstruct the point cloud and camera-pose for initial two-view reconstruction. We then perform depth-regularized optimization to register new images and triangulate the new points during incremental reconstruction. To comprehensively evaluate our approach, we introduce a new dataset (StudioSfM) consisting of 130 shots with 21K frames from 15 studio-produced videos that are manually annotated by a professional CG studio. We demonstrate that our approach: (a) significantly improves the quality of 3D reconstruction for our small-parallax setting, (b) does not cause any degradation for data with large-parallax, and (c) maintains the generalizability and scalability of geometry-based sparse SfM. Our dataset can be obtained at github.com/amazon-research/small-baseline-camera-tracking.

Talking head video generation aims to produce a synthetic human face video that contains the identity and pose information respectively from a given source image and a driving video. Existing works for this task heavily rely on 2D representations (e.g. appearance and motion) learned from the input images. However, dense 3D facial geometry (e.g. pixel-wise depth) is extremely important for this task as it is particularly beneficial for us to essentially generate accurate 3D face structures and distinguish noisy information from the possibly cluttered background. Nevertheless, dense 3D geometry annotations are prohibitively costly for videos and are typically not available for this video generation task. In this paper, we introduce a self-supervised face-depth learning method to automatically recover dense 3D facial geometry (i.e. depth) from the face videos without the requirement of any expensive 3D annotation data. Based on the learned dense depth maps, we further propose to leverage them to estimate sparse facial keypoints that capture the critical movement of the human head. In a more dense way, the depth is also utilized to learn 3D-aware cross-modal (i.e. appearance and depth) attention to guide the generation of motion fields for warping source image representations. All these contributions compose a novel depth-aware generative adversarial network (DaGAN) for talking head generation. Extensive experiments conducted demonstrate that our proposed method can generate highly realistic faces, and achieve significant results on the unseen human faces.

It is well known that the passive stereo system cannot adapt well to weak texture objects, e.g., white walls. However, these weak texture targets are very common in indoor environments. In this paper, we present a novel stereo system, which consists of two cameras (an RGB camera and an IR camera) and an IR speckle projector. The RGB camera is used both for depth estimation and texture acquisition. The IR camera and the speckle projector can form a monocular structured-light (MSL) subsystem, while the two cameras can form a binocular stereo subsystem. The depth map generated by the MSL subsystem can provide external guidance for the stereo matching networks, which can improve the matching accuracy significantly. In order to verify the effectiveness of the proposed system, we built a prototype and collected a test dataset in indoor scenes. The evaluation results show that the Bad 2.0 error of the proposed system is 28.2% of the passive stereo system when the network RAFT is used. The dataset and trained models are available at https://github.com/YuhuaXu/MonoStereoFusion.

Local density of point clouds is crucial for representing local details, but has been overlooked by existing point cloud compression methods. To address this, we propose a novel deep point cloud compression method that preserves local density information. Our method works in an auto-encoder fashion: the encoder downsamples the points and learns point-wise features, while the decoder upsamples the points using these features. Specifically, we propose to encode local geometry and density with three embeddings: density embedding, local position embedding and ancestor embedding. During the decoding, we explicitly predict the upsampling factor for each point, and the directions and scales of the upsampled points. To mitigate the clustered points issue in existing methods, we design a novel sub-point convolution layer, and an upsampling block with adaptive scale. Furthermore, our method can also compress point-wise attributes, such as normal. Extensive qualitative and quantitative results on SemanticKITTI and ShapeNet demonstrate that our method achieves the state-of-the-art rate-distortion trade-off.