Fall 2026 GRASP Seminar - Huaizu Jiang, Northeastern University
Huaizu Jiang, a researcher at Northeastern University, discusses how learned 3D models and geometric principles can work together, with the GRASP Lab seminar audience.
In this talk
- 2:36
Where geometry fits in learned 3D vision
Jiang frames geometric knowledge in three layers: matching raw observations, algorithms such as triangulation and PnP, and exact constraints such as perspective projection. Learned models are increasingly handling raw data and replacing some geometric algorithms, while exact constraints remain available when needed.
- 5:23
Relighting in a learned scene representation
Jiang describes a relighting method that encodes a scene as latent tokens, then modifies them using tokens for light sources represented by rays. The model has no explicit 3D scene or material model, yet it produces moving shadows, handles point and area lights, and generalizes from synthetic training images to real objects. Keeping edits in latent space also slows quality loss over repeated lighting changes compared with pixel-space methods.
- 22:08
One correspondence model for images and point clouds
UniCorn is designed to find matches across 2D image pairs, image-to-point-cloud pairs, and 3D point-cloud pairs. Its transformer separates appearance and position streams so matching layers can be stacked and refined. Synthetic point clouds made from image-derived geometry improve results, especially for the data-scarce 2D-to-3D task; joint training helps that task but can hurt 2D matching.
- 35:00
Efficient SLAM on an edge device
Jiang's SLAM system combines a lightweight network for optical flow, confidence, and stereo depth with geometric modules for camera tracking, dynamic-point rejection, and loop closure. On a Jetson Nano, it uses under 1 GB of memory and runs at 18 frames per second. The system keeps mapping accuracy comparable to other methods while limiting optimization to camera poses.
- 45:14
Open questions for 3D models
Jiang raises open questions about supervision for 3D foundation models and how much geometry to use when inference budgets vary. He argues that geometric methods can be a useful fallback on constrained devices, while tasks involving dynamic objects and affordances require more than geometry alone. Human demonstrations also offer a setting where geometric priors for hands and bodies can help transfer skills to robots.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.