CVPR 2021 Highlights Reveal Breakthroughs in 3D Vision and Privacy
Researchers at CVPR 2021 unveil innovative methods for 3D point cloud generation, human pose estimation, and privacy-preserving image features. These papers push the boundaries of computer vision across medical, audio-visual, and autonomous navigation domains.
Researchers at CVPR 2021 introduce powerful new techniques for understanding and generating 3D data. One paper leverages diffusion probabilistic models to generate highly detailed 3D point clouds, while another extracts skeletal representations directly from point clouds using a framework called Point2Skeleton. Additionally, the SCANimate paper presents a weakly supervised method for creating realistic skinned clothed avatar networks.
Human pose estimation and behavioral analysis see significant advancements in this latest research. PoseAug offers a differentiable framework that artificially augments 3D human poses to improve training data efficiency, while another study closely examines how self-contact impacts human pose accuracy. Furthermore, a concept called Task Programming enables machines to learn data-efficient behavior representations, bridging the gap between raw visual data and complex actions.
Beyond 3D modeling, the presented papers tackle critical challenges in security, healthcare, and multi-modal learning. A collaborative team from ETH Zurich and Microsoft develops adversarial affine subspace embeddings to extract image features without compromising user privacy. In the medical field, researchers propose a new way to calibrate image segmentation by modeling multi-rater agreement, and other innovations include robust image style transfer, audio-visual instance discrimination, and a binary temporal geofence to improve autonomous navigation safety.