Research map / 3D vision / Depth and stereo
Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation
Friedhelm Hamann, Ziyun Wang, Ioannis Asmanis, Kenneth Chaney, Guillermo Gallego, Kostas Daniilidis
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune a synthetically pretrained BFlow with the self-supervised contrast loss with motion prior | BFlow · EVIMO2 | TEPE: 8.62916088104248 -> 6.14132022857666 | Improved |
| self-supervised fine-tuning vs supervised in-domain training | BFlow · EVIMO2 | TEPE: 6.14132022857666 vs 3.38006925582885 | Hurt |
| use K=32 nearest neighbor trajectories instead of 1 (ablation) | Ours (U-Net) · DSEC | EPE: 4.576 -> 3.2 | Improved |
| use K=32 instead of 8 neighbor trajectories (ablation) | Ours (U-Net) · DSEC | EPE: 3.512 -> 3.2 | Improved |
| use randomized reference time instead of three fixed reference times (ablation) | Ours (U-Net) · DSEC | EPE: 4.46 -> 3.2 (single reference time 7.264) | Improved |
| use non-linear (higher-degree or learned) motion prior on DSEC (ablation) | Ours (U-Net) · DSEC | EPE: 3.2 vs 3.22 (learned) and 3.28 (polynomial) | No clear effect |
| use Bezier curves (N_c=10) instead of polynomial or learned basis (ablation) | Ours (BFlow) · EVIMO2 | TEPE: 6.77561473846435 / 7.45736408233642 -> 6.14132022857666 | Improved |
| use non-linear Bezier prior instead of linear polynomial (N_c=1) on EVIMO2 (ablation) | Ours (BFlow) · EVIMO2 | TEPE: 7.97 -> 6.14 | Improved |
| use self-supervised contrast loss on night sequence zurich_city_14 with flickering lights | Ours (U-Net) · DSEC zurich_city_14_c | EPE: 1.78 vs E-RAFT 0.713 | Hurt |
| replace recurrent or RAFT-style networks with a single-pass U-Net trained with the loss | Ours (U-Net) · DSEC | inference time (ms): 40.1 / 46.331 -> 7.273 | Improved |
| Ours vs test-time optimization method on MVSEC | Ours (U-Net) · MVSEC | EPE: 0.45 vs Shiba 0.42 (indoor_flying1); 0.71 vs 0.60 (indoor_flying2) | Mixed |
Closest papers by meaning
- Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow (ICCV 2023 · similarity 0.91)
- Secrets of Event-Based Optical Flow, Depth and Ego-Motion Estimation by Contrast Maximization (TPAMI 2024 · similarity 0.90)
- From Contrast to Consistency: Rethinking Event-based Continuous-Time Optical Flow Estimation (CVPR 2026 · similarity 0.89)
- Dense Continuous-Time Optical Flow From Event Cameras (TPAMI 2024 · similarity 0.89)
- TimeTracker: Event-based Continuous Point Tracking for Video Frame Interpolation with Non-linear Motion (CVPR 2025 · similarity 0.86)
- FlowTrack: Revisiting Optical Flow for Long-Range Dense Tracking (CVPR 2024 · similarity 0.86)
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.
