Research map / 3D vision / Depth and stereo
DERD-Net: Learning Depth from Event-based Ray Densities
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 DERD-Net, a tiny 3D-conv + GRU network that predicts pixel-wise depth from local Sub-DSI patches of event ray-density volumes, in monocular and stereo event setups.
Problem
Event-camera depth estimation from disparity space images (DSIs) typically uses argmax over depth, which ignores local ray-intersection patterns and yields few, noisy depth estimates; deep learning on raw asynchronous events overfits and is costly.
Approach
Process small normalized local Sub-DSI patches around selected pixels with a tiny 3D-convolution + GRU network (~70k parameters) that regresses depth, run in parallel per pixel, making cost independent of camera resolution and improving density and accuracy over argmax-based MC-EMVS.
Key results
- On MVSEC stereo with F_orig, DERD-Net reduces mean error to 11.69 cm vs 20.07 cm for MC-EMVS (about 42% lower MAE on the same pixels).
- Monocular DERD-Net reaches 23.68 cm mean error vs 33.78 cm for monocular EMVS, and comparable to stereo SOTA.
- On DSEC Zurich_City_04_a, multi-pixel DERD-Net with F_orig reduces mean error by 55% versus MC-EMVS with morphological filter, with a 1.72x higher point count.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DERD-Net instead of EMVS for monocular DSIs with F_orig | Monocular depth estimation · MVSEC indoor_flying | Mean Err [cm]: 33.78 -> 23.68 | Helped |
| Use DERD-Net instead of MC-EMVS for stereo DSIs with F_orig | Stereo depth estimation · MVSEC indoor_flying | Mean Err [cm]: 20.07 -> 11.69 | Helped |
| Use DERD-Net instead of MC-EMVS for stereo DSIs with F_denser | Stereo depth estimation · MVSEC indoor_flying | Mean Err [cm]: 28.38 -> 15.24 | Helped |
| Use DERD-Net instead of MC-EMVS for stereo DSIs with F_orig | Stereo depth estimation · DSEC Zurich_City_04_a | Mean Err [m]: 3.27 -> 1.61 | Helped |
| Use DERD-Net instead of MC-EMVS for stereo DSIs with F_denser | Stereo depth estimation · DSEC Zurich_City_04_a | Mean Err [m]: 4.76 -> 1.80 | Helped |
| Use multi-pixel output instead of single-pixel output (ablation) | DERD-Net · MVSEC indoor_flying, stereo + F_orig | #Points [million]: 0.79 -> 4.32 | Helped |
| Use multi-pixel output instead of single-pixel output (ablation) | DERD-Net · DSEC Zurich_City_04_a, stereo + F_orig | Mean Err [m]: 1.61 -> 1.59 | Helped |
| Remove ensemble learning (ablation) | DERD-Net · MVSEC indoor_flying, stereo + F_orig | Mean Err [cm]: 12.0 -> 11.69 | Hurt |
| Increase Sub-DSI frame size from 7 × 7 to 9 × 9 (ablation) | DERD-Net · MVSEC indoor_flying_1, stereo + F_orig | Mean Err [cm]: 12.13 -> 11.58 | Helped |
| Increase Sub-DSI frame size from 7 × 7 to 9 × 9 (ablation) | DERD-Net · MVSEC indoor_flying_1, stereo + F_denser | Mean Err [cm]: 15.41 -> 14.59 | Helped |
| Horizontally reflect the DSI (ablation) | DERD-Net · MVSEC indoor_flying_1, stereo + F_orig | Mean Err [cm]: 11.69 -> 14.26 | Hurt |
| Vertically reflect the DSI (ablation) | DERD-Net · MVSEC indoor_flying_1, stereo + F_orig | Mean Err [cm]: 11.69 -> 12.81 | Hurt |
Surprising findings
- Reprojection-based pose noise from DERD-Net's own depth only slightly degrades results, with no metric worsening by more than 13%.
- Performance on MVSEC indoor_flying 1 with ES-PTAM poses improved slightly on bad-pix despite noisy poses.
Limitations
- Sub-DSI size trades accuracy against parameter count and compute; larger patches help but increase cost.
- Evaluated on only one DSEC sequence and MVSEC indoor flying; training per sequence split with cross-validation.
- Depends on accurate camera poses for DSI construction; DSI creation dominates runtime.
Experimental setup
Datasets: MVSEC indoor_flying DSEC Zurich_City_04_a
Benchmarks: MVSEC indoor_flying DSEC Zurich_City_04_a
Metrics: mean error median error bad-pix SILog error AErrR log RMSE delta accuracy number of points
Baselines: EMVS MC-EMVS ESVO SGM GTS ESVO2 DDES EIT-Net StereoSpike SCS-Net
Builds on: EMVS MC-EMVS FireNet RAFT-Stereo
Closest papers by meaning
- Deep Event Stereo Leveraged by Event-to-Image Translation (AAAI 2021 · similarity 0.91)
- Stereo Depth From Events Cameras: Concentrate and Focus on the Future (CVPR 2022 · similarity 0.91)
- Event-Driven Dynamic Scene Depth Completion (NeurIPS 2025 · similarity 0.89)
- Event-Image Fusion Stereo Using Cross-Modality Feature Propagation (AAAI 2022 · similarity 0.88)
- Active Event-based Stereo Vision (CVPR 2025 · similarity 0.88)
- Zero-Shot Event-Intensity Asymmetric Stereo via Visual Prompting from Image Domain (NeurIPS 2024 · similarity 0.87)
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.