My broad research interest is 3D reconstruction in terms of 3d geometry, ego‑motion and object motion. A particular focus is robust application on arbitrary real‑world data.
We introduce ADE-OoD, a novel benchmark for out-of-distribution (OoD) detection for semantic segmentation on general natural images, and DOoD, a diffusion-based method that detects OoD samples by perturbing them and checking the directional error of the estimated score.
We generate 3D shape and appearance of objects via diffusion on neural point clouds (points with a 3D position and a learned feature). As the point positions represent the coarse shape and the features represent the appearance, shape and appearance can be generated separately.
We introduce a benchmark for robust zero-shot multi-view depth estimation and evaluate many recent models. We present a new model that works more robustly across data from different domains and can serve as a baseline for future evaluations on the benchmark.
We propose SF2SE3, an approach estimate a moving object segmentation and corresponding SE(3) motions from optical flow and depth. It works by iteratively sampling motion proposals and selecting the best ones with respect to a maximum coverage formulation.
We train a network for disparity estimation in a semi-supervised fashion on labeled synthetic data and unlabeled real data. For supervision on the unlabeled data, we explore a deep feature reconstruction loss, instead of the commonly used photometric loss.
The template for this website comes from Jon Barron. Thanks!
