Ranked 2nd 🥈 in Robotic Vision Scene Understanding Challenge 2023

Ranked 2nd at the Robotic Vision Scene Understanding Challenge, hosted at the Embodied-AI Workshop of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2023, which benchmarked methods for joint semantic and geometric scene understanding in embodied settings. The challenge evaluated two core problems: object-centric Semantic SLAM (building an object-level semantic map from RGB-D and odometry) and scene change detection (identifying additions/removals between mapping runs, including condition shifts such as day–night). It was particularly demanding due to evaluation across multiple regimes for e.g. passive vs. active perception and ideal vs. dead-reckoning localization (with drift), under realistic robot simulation pipelines.