ICCV 2025 · Highlight Paper
A Linear N-Point Solver for Asynchronous Tracks
Nov 2024 – Mar 2025 · ShanghaiTech Mobile Perception Lab
A unified linear solver for recovering structure and continuous motion from point tracks captured at arbitrary timestamps.
Motivation
Classical five- and eight-point algorithms assume synchronized views. That assumption breaks for rolling-shutter cameras, event cameras, and mixed sensing systems, where each correspondence may carry a different timestamp.
Formulation
The project models motion with first-order dynamics and a constant-velocity assumption. This yields a linear point-incidence relation that jointly recovers linear velocity and 3D structure from tracks observed at arbitrary times and across an arbitrary collection of views.
Sensor coverage
The same formulation handles global-shutter images, rolling-shutter measurements, event-camera tracks, and correspondences combined from collocated sensors. Its linear structure also makes degeneracies and solution multiplicities easier to characterize.
Results
Experiments on simulated and real data showed consistent gains over recent modality-specific approaches. The implementation reached 63 μs CPU runtime and reduced velocity error by 10–30% in the reported comparisons.
My contribution
I developed real-time LED-matrix detection and contributed to encoding and tracking modules. I also integrated asynchronous feature trackers, including ArcStar and RATE, to supply accurate temporally varying point tracks to the solver.