Research map / 3D vision / Depth and stereo
Event-based Structure-from-Orbit
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes event-based structure-from-orbit (eSfO), jointly recovering sparse 3D structure and orbital/spin parameters of a spinning object observed by a static event camera, via a new clustering-based event feature tracker (ETC) and a factor-graph orbit back-end, plus a new TOPSPIN turntable event dataset.
Problem
Event-only 3D reconstruction of a fast spinning object observed by a static event camera, recovering spin rate, rotation axis and sparse shape; existing event feature trackers fail under periodic self-occlusion from spinning motion.
Approach
Equate a static event camera observing a spinning object with an orbiting camera on a circular path; track features via spatio-temporal HDBSCAN clustering of corner events (ETC) and refine a constrained orbit model with a factor-graph reprojection optimization initialized from COLMAP and an FFT frequency estimate.
Key results
- On the Event Camera Dataset, ETC achieved lower reprojection RMSE than HASTE, Metavision and eCDT on shapes and boxes at most thresholds (e.g., shapes 1.70 vs HASTE 2.37 at 7px threshold).
- On TOPSPIN spinning-object scenes, HASTE tracks yielded a mean COLMAP reprojection error of 348.48 px while ETC achieved 2.39 px.
- Fourier-transform initialization of spin rate is refined by eSfO optimization toward turntable ground truth frequency (Fig. 8a, no numeric values in text).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ETC instead of HASTE for tracking spinning-object events | event feature tracking · hubble-diagonal-fast | COLMAP reprojection error (pixel): 348.48 -> 2.39 | Helped |
| Use ETC instead of Metavision | event feature tracking · Event Camera Dataset: shapes, poster, and boxes | feature age (seconds): shapes: 0.007 -> 0.332; poster: 0.005 -> 0.499; boxes: 0.005 -> 0.525 | Helped |
| Use ETC instead of Metavision | event feature tracking · Event Camera Dataset: shapes, poster, and boxes | RMSE (pixel), threshold 3 pixels: shapes: 0.17 -> 0.88; poster: 0.13 -> 0.95; boxes: 0.13 -> 0.88 | Hurt |
| Constrain the camera trajectory to the eSfO circular-orbit model rather than unconstrained COLMAP poses | eSfO · TOPSPIN | reprojection error | Hurt |
| Refine the Fourier Transform frequency initialization with eSfO optimization | eSfO frequency estimation · TOPSPIN | frequency estimation accuracy | Helped |
Surprising findings
- eSfO optimization yields higher average reprojection error than the unconstrained COLMAP reconstruction, but produces a better orbital trajectory due to the circular constraint.
- Metavision tracker produced extremely short feature tracks (mean age around 0.005-0.007 s), limiting usefulness.
Limitations
- Feature age and tracking quality depend on rotational speed, since faster spin causes self-occlusion sooner.
- Window size delta-t for track extraction trades off tracking accuracy against number of views.
- Reconstruction is sparse and the orbit model is restricted to single-axis circular motion.
Experimental setup
Datasets: Event Camera Dataset TOPSPIN
Benchmarks: Event Camera Dataset TOPSPIN
Metrics: RMSE reprojection error (pixels) feature age (seconds) estimated frequency vs ground truth point cloud to CAD registration error
Baselines: HASTE Metavision feature tracker eCDT eCDT (w/o HT)
Builds on: eFAST HDBSCAN COLMAP HASTE
Closest papers by meaning
- EventPillars: Pillar-based Efficient Representations for Event Data (AAAI 2025 · similarity 0.86)
- ETAP: Event-based Tracking of Any Point (CVPR 2025 · similarity 0.84)
- IncEventGS: Pose-Free Gaussian Splatting from a Single Event Camera (CVPR 2025 · similarity 0.84)
- EF-3DGS: Event-Aided Free-Trajectory 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.84)
- FlexEvent: Towards Flexible Event-Frame Object Detection at Varying Operational Frequencies (NeurIPS 2025 · similarity 0.83)
- TimeTracker: Event-based Continuous Point Tracking for Video Frame Interpolation with Non-linear Motion (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.