Featured

An end-to-end differentiable model for relative camera pose estimation from RGB-D frame pairs. The network learns self-supervised confidence scores for predicted correspondences, which are then aligned by a differentiable weighted Procrustes solver. Evaluated on ScanNet with strong results in wide-baseline and occlusion-heavy regimes.

We present an end-to-end learnable, differentiable method for pairwise relative pose registration of RGB-D frames. Our method is robust to big camera motions thanks to a self-supervised weighting of the predicted correspondences between the frames. Given a pair of frames, our method estimates matches of points and their visibility score. A self-supervised model predicts a confidence weight for visible matches. Finally, visible matches and their weight are fed into a differentiable weighted Procrustes aligner which estimates the rigid transformation between the input frames.

Implementation of the paper “Kinect Fusion: Real-Time Dense Surface Mapping and Tracking”

Worked on the SLAM pipeline of an autonomous driving vehicle using ROS, C++, PCL, Ceres Solver, and Google Cartographer. Developed a ROS package that turns SLAM point clouds into a cost map encoding both occupancy and surface preference (e.g., asphalt over grass) for path planning.

Extended Eisenberger et al.’s divergence-free shape interpolation by introducing time-dependent vector fields, lifting the limitation that different parts of a shape cannot move in conflicting directions through the same point in space. Correspondence and matching are jointly solved over the full sequence during optimization.

Built a C++ library and an efficient CNF encoding for solving Pseudo-Boolean Optimization problems via off-the-shelf SAT solvers, with linear and binary minimization strategies and configurable timeout policies for time-bounded solving.