Image features with differential privacy guarantee for visual localization and structure-from-motion.

Revisiting NOCS-based 3D object detection and localization in driving scenes with categorical object NeRF.

Test-time domain adapation for 3D semantic segmentation leveraging image-Lidar cross-modal consistency.

Geometric transformation creates dense occlusion-aware semantic map from compact parametric annotation, facilitating road layout estimation.

A cross-domain and cross-modal (geometry and appearance) constrastive learning framework for video domain adpatation.

A principled way to fuse the geometric camera pose estimation with CNN predictions, via learning geometry-guided uncertainty, driven by a self-attention graph neural network.

Geometric SLAM and self-supervised monocular CNN-depth learning can benefit each other.

Joint rolling shutter image stitching and distortion removal, by deriving a new rolling-shutter-aware homography and a minimal 5-point solver.

Learning parametric road layout from non-parametric semantic 3D reconstruction obtained by Structure-from-Motion.

Theoretical degeneracy on SfM with a rolling-shutter camera and leveraging data-driven priors through a network that learns camera motion and scene structure to undistort a single rolling shutter image.

Theoretical degeneracy on radial distortion self-calibration in forward motion and a network to learn radial distortion parameters and camera intrinsics for SLAM.

A baseline-insensitive bilinear objective function for translation averaging in global SfM. Theoretically revealing the underlying subtle difference that leads to the performance gap between two convex methods, LUD and Shapefit/kick.

Develop a rolling-shutter-aware differential SfM method for depth and motion recovery, which is further leveraged to remove rolling shutter distortion.