A fully online system to effectively integrate dense CLIP features with Gaussian Splatting. High-resolution dense CLIP embedding and online compressor learning modules are introduced to serve dense language mapping at realtime (40+ FPS) while retaining open-vocabulary capability for flexible query-based human-machine interaction.

Depth Any Camera (DAC) is a training framework for metric depth estimation that enables zero-shot generalization across cameras with diverse fields of view—including fisheye and 360° images. Tired of collecting new data for every camera setup? DAC maximizes the utility of existing 3D datasets, making them applicable to a wide range of camera types without the need for retraining.

A monocular object reconstruction framework effectively integrating object pose estimation and NeRF-based reconstruction. A novel camera-invariant pose estimation module is introduced to resolve depth-scale ambiguity and enhance cross-domain generalization.

A mathematical framework to prove that the dice loss leads to superior noise-robustness and model convergence for large objects compared to regression losses. A flexible monocular 3D detection pipeline integrated with bird-eye view segmentation.

A neural reconstruction method enabling the completion the occluded surfaces from large 3D scene reconstrucion. A milestone in automating the creation of interactable digital twins from real world.

A keypoint-based object-level SLAM framework that can provide globally consistent 6DoF pose estimates for symmetric and asymmetric objects.

The first vision transformer approach to handle 360 monocular depth estimation with spherical distortion. Novel designs include tangent-image coordinate embedding and geometry-aware feature fusion.

A pioneer work in predicting 3D lanes from a single image with high generalization to novel scenes. A 3D lane synthetic dataset is introduced.

Numerically robust image filters and symbolic curve models are introduced to precisely estimate differential geometry attributes associated with image edges, including localization, orientation, curvature, edge topology.

A joint model of learned part-based appearance and parametric shape representation to precisely estimate the highly articulated poses of multiple laboratory animals.

This study investigates the relative diagnosticity and the optimal combination of multiple cues (we consider luminance, color, motion and binocular disparity) for boundary detection in natural scenes. A multi-cue boundary dataset is introduced to facilitate the study.