Research map / 3D vision / Depth and stereo
Active Event-based Stereo Vision
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 active event-based stereo vision, combining binocular event cameras with an infrared pattern projector, builds a real and a simulated dataset, and proposes ActiveEventNet, a lightweight stereo matching network.
Problem
Frame-based active stereo cameras are limited by depth frame rates and motion blur in high-speed scenes, while passive event-based stereo struggles in texture-less and low-light scenes; existing event-based stereo models are also computationally heavy.
Approach
Integrate binocular DAVIS346 event cameras with an 850 nm infrared pattern projector so structured light adds texture and generates events even in static or dark scenes, and use a MobileNet-based network with a dynamic channel-interaction cost volume for fast dense disparity estimation.
Key results
- ActiveEventNet reaches EPE 1.995, RMSE 7.821, D1-all 0.083 on the real-world Active-Event-Stereo dataset with 2.23M params and 6.5 ms runtime, versus Unimatch at EPE 1.902 and 87.6 ms.
- On the synthetic RealSense-Event-Sim, ActiveEventNet achieves EPE 1.223 and D1-all 0.070, better than PSMNet, DeepPruner-Fast, AANet, and DDES.
- Structured light lowers EPE from 2.614 to 1.993 on static and slow-motion sequences compared with passive event input; the prototype runs at up to 150 FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ActiveEventNet instead of RGB-frame stereo | Active stereo on the Active-Event-Stereo dataset · Active-Event-Stereo average | EPE: 2.238 -> 1.995 | Helped |
| Use ActiveEventNet instead of Unimatch | stereo matching on Active-Event-Stereo · Active-Event-Stereo test | Runtime (ms): 87.6 -> 6.5 | Helped |
| Use ActiveEventNet instead of Unimatch | stereo matching on Active-Event-Stereo · Active-Event-Stereo test | EPE: 1.902 -> 1.995 | Hurt |
| Use ActiveEventNet instead of Unimatch | stereo matching on RealSense-Event-Sim · RealSense-Event-Sim test | EPE: 1.108 -> 1.223 | Hurt |
| Add MobileNet blocks while retaining concatenation cost volume (ablation) | ActiveEventNet · Active-Event-Stereo test | Runtime (ms): 33.8 -> 6.2 | Helped |
| Add MobileNet blocks while retaining concatenation cost volume (ablation) | ActiveEventNet · Active-Event-Stereo test | EPE: 2.124 -> 1.968 | Helped |
| Add dynamic interaction cost volume while retaining standard convolutions (ablation) | ActiveEventNet · Active-Event-Stereo test | EPE: 2.124 -> 2.21 | Hurt |
| Use MobileNet v2 blocks in both feature extraction and encoder-decoder (ablation) | ActiveEventNet · Active-Event-Stereo test | Runtime (ms): 35.7 -> 6.5 | Helped |
| Replace concatenation cost volume with dynamic interaction (ablation) | ActiveEventNet · Active-Event-Stereo test | EPE: 2.527 -> 1.993 | Helped |
| Replace correlation cost volume with dynamic interaction (ablation) | ActiveEventNet · Active-Event-Stereo test | EPE: 2.316 -> 1.993 | Helped |
| Add structured light (ablation) | binocular event-based stereo vision · Active-Event-Stereo test | EPE: 2.614 -> 1.993 | Helped |
Surprising findings
- Event-based active stereo with sparse events outperforms dense RGB stereo in most real-world sequences, including several under normal light.
- Replacing standard convolutions with MobileNet blocks gave nearly 6x speedup with only a small accuracy drop.
Limitations
- Real-world labels come from a RealSense D455 at 15 Hz, which is itself a frame-based sensor, so ground truth is limited in rate and accuracy.
- Accuracy gains from dynamic interaction and MobileNet blocks are small; the best absolute accuracy on the real dataset is from Unimatch, which is much slower.
Experimental setup
Datasets: Active-Event-Stereo (real-world, newly built) RealSense-Event-Sim (synthetic, newly built)
Benchmarks: Active-Event-Stereo RealSense-Event-Sim
Metrics: EPE RMSE D1-all >1px >2px >3px #Params Runtime (ms)
Baselines: SGM PSMNet DeepPruner-Fast AANet Unimatch DDES RealSense D455
Builds on: PSMNet AANet DDES MobileNet v2 E2VID V2E
Closest papers by meaning
- A Low Power, High Throughput, Fully Event-Based Stereo System (CVPR 2018 · similarity 0.91)
- Stereo Depth From Events Cameras: Concentrate and Focus on the Future (CVPR 2022 · similarity 0.90)
- DERD-Net: Learning Depth from Event-based Ray Densities (NeurIPS 2025 · similarity 0.88)
- Zero-Shot Event-Intensity Asymmetric Stereo via Visual Prompting from Image Domain (NeurIPS 2024 · similarity 0.87)
- Event-Driven Dynamic Scene Depth Completion (NeurIPS 2025 · similarity 0.87)
- EventPS: Real-Time Photometric Stereo Using an Event Camera (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.