A unified feed-forward transformer for unseen object pose estimation that jointly predicts depth, point maps, camera parameters, and NOCS to recover both absolute and relative poses.

Our paper HouseCat-TRICKY has been accepted to ICCVW 2025. The benchmark and evaluation focus on challenging transparent and reflective object categories.

Texture2LoD3 proposes leveraging ubiquitous street-level images and low-level building models for accurate ortho-texturing (left): Enabling accurate semantic segmentation (center) and facade-rich LoD3 reconstruction (right).

A semantic shape reconstruction module that recovers complete object geometry from partial observations with a global context-enhanced feature fusion mechanism that leverages category-level semantic and shape priors for robust pose prediction

A novel method to achieve Gaussian splatting from an un-posed image pair in dynamic environments.

A framework to annotate accurate and dense 3d indoor scenes with a benchmark on novel view synthesis and SLAM

A knowledge-based Programming by Demonstration (kb-PbD) paradigm to facilitate robot programming in small and medium-sized enterprises (SMEs).