Research map / 3D vision / Depth and stereo
Zero-Shot Event-Intensity Asymmetric Stereo via Visual Prompting from Image Domain
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 ZEST, a training-free zero-shot event-intensity asymmetric stereo method that aligns frames and events via a temporal-difference/temporal-integral visual prompt, feeds off-the-shelf image stereo models, and refines disparity with monocular depth from Depth Anything or MiDaS.
Problem
Event-intensity asymmetric stereo matching suffers from overfitting and poor generalization because event datasets are small and lack scene diversity, and modality gaps between events and frames hinder use of image-domain stereo models.
Approach
Align frames and events into a shared intermediate representation (temporal difference of frames matched to temporal integral of events, with exposure-time modeling) so frozen image-domain stereo models can be applied zero-shot, then refine disparity by fitting a locally linear scale/shift of monocular depth guided by event-density confidence.
Key results
- Ours-DS-DA achieves 3PE 15.05% on DSEC (all sequences) versus 86.96% for DAEI trained on MVSEC, and Ours-CR-DA reaches 9.84% 3PE.
- Ours-CR-DA on MVSEC obtains EPE 5.22 versus 10.57 for HSM, but DAEI (1.08 EPE) remains better on its in-domain split.
- On DSEC, the in-dataset fine-tuned DAEI† reaches lower EPE (1.93) than ZEST variants, which the authors attribute to in-dataset training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ZEST with CREStereo and Depth Anything rather than the cross-dataset baselines | Event-intensity stereo on DSEC · DSEC All | EPE: 10.17 -> 3.99 | Helped |
| Use ZEST with DynamicStereo and Depth Anything rather than the cross-dataset baselines | Event-intensity stereo on DSEC · DSEC All | EPE: 13.83 -> 2.99 | Helped |
| Use ZEST with CREStereo and Depth Anything rather than CR-E2VID | Event-intensity stereo · M3ED | EPE: 2.10 -> 2.06 | Helped |
| Add the proposed temporal representation alignment instead of unaligned E2VID reconstruction (ablation) | DynamicStereo on interlaken_00_c · interlaken_00_c (DSEC) | EPE: 43.02 -> 1.49 | Helped |
| Add the proposed temporal representation alignment instead of v2e-converted left events (ablation) | DynamicStereo on interlaken_00_c · interlaken_00_c (DSEC) | EPE: 13.90 -> 1.49 | Helped |
| Replace spatial gradients and spatial event integrals with temporal difference and temporal event integral representations (ablation) | DynamicStereo on interlaken_00_c · interlaken_00_c (DSEC) | EPE: 19.01 -> 1.49 | Helped |
| Add DA_Large refinement to spatial-gradient/spatial-integral stereo results (ablation) | DynamicStereo on interlaken_00_c · interlaken_00_c (DSEC) | EPE: 19.01 -> 19.09 | Hurt |
| Add monocular cue-guided refinement to ZEST (ablation) | DynamicStereo on interlaken_00_c · interlaken_00_c (DSEC) | EPE: 1.49 -> 1.41 | Helped |
| Replace DA_Large with DA_Base for refinement (ablation) | ZEST with DynamicStereo · interlaken_00_c (DSEC) | EPE: 1.41 -> 1.39 | Helped |
| Replace DA_Large with DA_Small for refinement (ablation) | ZEST with DynamicStereo · interlaken_00_c (DSEC) | EPE: 1.41 -> 1.42 | No clear effect |
Surprising findings
- The in-dataset fine-tuned DAEI† beats most ZEST variants on EPE and RMSE on DSEC, though ZEST variants win on 3PE.
Limitations
- Fixed (non-learned) representation alignment may not capture the full event-frame modality gap.
- Noisy or sparse events degrade visual prompts and refinement, causing failure cases.
- Large off-the-shelf models and per-image optimization make inference slow (under 2 FPS for CR-based variants).
Experimental setup
Datasets: DSEC MVSEC M3ED
Benchmarks: DSEC MVSEC M3ED
Metrics: EPE RMSE 3PE 2PE 1PE
Baselines: SHEF HSM DAEI PSMNet-ETNet CR-E2VID DS-E2VID CFF-v2e CR-ETNet DS-ETNet PSMNet-E2VID
Builds on: CREStereo DynamicStereo Depth Anything MiDaS E2VID
Closest papers by meaning
- Stereo Anywhere: Robust Zero-Shot Deep Stereo Matching Even Where Either Stereo or Mono Fail (CVPR 2025 · similarity 0.89)
- Active Event-based Stereo Vision (CVPR 2025 · similarity 0.87)
- DERD-Net: Learning Depth from Event-based Ray Densities (NeurIPS 2025 · similarity 0.87)
- FoundationStereo: Zero-Shot Stereo Matching (CVPR 2025 · similarity 0.86)
- MVSAnywhere: Zero-Shot Multi-View Stereo (CVPR 2025 · similarity 0.86)
- PPMStereo: Pick-and-Play Memory Construction for Consistent Dynamic Stereo Matching (NeurIPS 2025 · similarity 0.85)
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.