Research map / 3D vision / Depth and stereo
EventPS: Real-Time Photometric Stereo Using an Event Camera
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.
Introduces EventPS, a photometric stereo method that estimates surface normals from event-camera radiance changes under a rotating LED, with optimization (SVD) and deep-learning (PS-FCN/CNN-PS adapted) variants, reaching ~30 fps and ~26-40% of FramePS data rate.
Problem
Conventional frame-based photometric stereo needs many HDR images under sequential lighting, which is slow, bandwidth-heavy, and unsuitable for real-time or dynamic scenes.
Approach
Reformulates photometric stereo to use only event-triggered radiance changes under continuously rotating lighting: each pair of consecutive events yields a null space vector orthogonal to the surface normal, and normals are recovered by SVD or by feeding null-space-vector maps into PS-FCN and CNN-PS.
Key results
- On DiLiGenT-Ev, EventPS requires on average 25.9% (EventPS-OP), 39.8% (EventPS-FCN), and 26.6% (EventPS-CNN) of FramePS data rate for equal MAE.
- On real objects, average MAE is 18.8 (EventPS-OP), 14.7 (EventPS-FCN), and 17.6 (EventPS-CNN).
- EventPS-OP runs at around 30 fps in real-time normal estimation with GPU-accelerated pipeline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use event-camera observations instead of frame-image observations | EventPS versus FramePS counterparts · DiLiGenT-Ev | data rate: 25.86%; 39.82%; 26.61% of FramePS data rate | Helped |
| use event-camera observations instead of frame-image observations | EventPS versus FramePS counterparts · DiLiGenT-Ev | MAE | No clear effect |
| increase event-triggering noise std from 0.05 to 0.2 (ablation) | EventPS-OP · DiLiGenT-Ev | MAE: 23.6 -> 28.1 | Hurt |
| increase event-triggering noise std from 0.05 to 0.2 (ablation) | EventPS-FCN · DiLiGenT-Ev | MAE: 14.4 -> 15.7 | Hurt |
| increase event-triggering noise std from 0.05 to 0.2 (ablation) | EventPS-CNN · DiLiGenT-Ev | MAE: 16.9 -> 19.6 | Hurt |
Surprising findings
- Event-based normal estimation remains robust as event noise increases, with all three variants showing graceful MAE degradation.
- EventPS-CNN runs at only about 0.1 fps and EventPS-FCN at about 2 fps, far slower than EventPS-OP's over 1000 fps processing speed.
Limitations
- Circular scanning pattern leaves a blind area for high elevation-angle normals; hypotrochoid pattern is hard to implement mechanically.
- Event signal quality degrades as lighting scan speed increases due to event camera frequency response.
- Evaluated mainly under Lambertian-derived formulation with rotating mechanical LED; deep variants are slow (EventPS-CNN ~0.1 fps).
Experimental setup
Datasets: Blobby Sculpture DiLiGenT DiLiGenT-Ev
Benchmarks: DiLiGenT-Ev real dataset (5 fabricated objects)
Metrics: mean angular error (MAE) data rate
Baselines: TH28 PS-FCN CNN-PS
Builds on: PS-FCN CNN-PS ESIM
Closest papers by meaning
- EventPSR: Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera (CVPR 2025 · similarity 0.92)
- PS-EIP: Robust Photometric Stereo Based on Event Interval Profile (CVPR 2025 · similarity 0.91)
- Active Event-based Stereo Vision (CVPR 2025 · similarity 0.86)
- Differentiable Display Photometric Stereo (CVPR 2024 · similarity 0.85)
- Event-3DGS: Event-based 3D Reconstruction Using 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.85)
- Zero-Shot Event-Intensity Asymmetric Stereo via Visual Prompting from Image Domain (NeurIPS 2024 · 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.