Research map / 3D vision / Depth and stereo
Event-Driven Dynamic Scene Depth Completion
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 EventDC, an event-driven depth completion framework using event-modulated deformable alignment in the encoder and motion-mask-guided local filtering in the decoder, plus a new event-based depth completion benchmark with one real and two synthetic datasets.
Problem
Depth completion in dynamic scenes is degraded by ego-motion and object motion, which blur RGB and misalign LiDAR; existing methods have no event-based depth completion data or methods.
Approach
Use event features to predict offsets, weights and motion masks that modulate deformable convolutions for RGB-D alignment (encoder, EMA) and local depth refinement around moving objects (decoder, LDF), supervised by structure-aware and motion-aware losses.
Key results
- On EventDC-Real, EventDC reaches RMSE 574.0 mm and MAE 179.0 mm, beating the second-best method (Prompting) by 84.1 mm RMSE and 26.1 mm MAE.
- On EventDC-FullSyn, EventDC reduces RMSE by 53.5 mm and MAE by 19.4 mm over the second-best approach.
- Ablation on EventDC-Real: adding EMA and LDF together lowers RMSE from 638.3 (event+RGB+depth fusion) to 574.0 mm.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EventDC instead of FusionNet | EventDC on EventDC-Real · EventDC-Real | RMSE (mm): 658.1 -> 574.0 | Helped |
| Use EventDC instead of OGNI-DC | EventDC on EventDC-SemiSyn · EventDC-SemiSyn | RMSE: 832.0 -> 778.8 | Helped |
| Use EventDC instead of OGNI-DC | EventDC on EventDC-FullSyn · EventDC-FullSyn | RMSE: 673.7 -> 620.2 | Helped |
| Add RGB input to the depth-only baseline (ablation) | EventDC depth-only baseline on EventDC-Real · EventDC-Real | RMSE: 727.2 -> 685.9 | Helped |
| Add event input to the depth-only baseline (ablation) | EventDC depth-only baseline on EventDC-Real · EventDC-Real | RMSE: 727.2 -> 696.0 | Helped |
| Use depth, RGB, and event inputs rather than depth-only input (ablation) | EventDC on EventDC-Real · EventDC-Real | RMSE: 727.2 -> 638.3 | Helped |
| Add encoder dynamic convolution (ablation) | EventDC with depth, RGB, and event inputs on EventDC-Real · EventDC-Real | RMSE: 638.3 -> 628.8 | Helped |
| Add encoder EMA (ablation) | EventDC with depth, RGB, and event inputs on EventDC-Real · EventDC-Real | RMSE: 628.8 -> 602.8 | Helped |
| Add decoder LDF rather than decoder dynamic convolution (ablation) | EventDC with depth, RGB, and event inputs on EventDC-Real · EventDC-Real | RMSE: 630.6 -> 605.4 | Helped |
| Use both encoder EMA and decoder LDF (ablation) | EventDC with depth, RGB, and event inputs on EventDC-Real · EventDC-Real | RMSE: 683.3 -> 574.0 | Helped |
Limitations
- Relies on high-quality event data and precise sensor alignment that may not be available in all settings.
- EMA and LDF add computational cost, limiting deployment on resource-constrained devices.
- Real-world dataset scale and diversity are limited.
Experimental setup
Datasets: EventDC-Real EventDC-SemiSyn EventDC-FullSyn KITTI CARLA VID2E
Benchmarks: EventDC-Real EventDC-SemiSyn EventDC-FullSyn
Metrics: RMSE MAE REL delta 1.05 delta 1.10 delta 1.15
Baselines: CSPN S2D FusionNet RigNet DySPN Prompting OGNI-DC SigNet LPNet
Builds on: DCNv2 DySPN RigNet CSPN
Closest papers by meaning
- DERD-Net: Learning Depth from Event-based Ray Densities (NeurIPS 2025 · similarity 0.89)
- Bring Event into RGB and LiDAR: Hierarchical Visual-Motion Fusion for Scene Flow (CVPR 2024 · similarity 0.87)
- Active Event-based Stereo Vision (CVPR 2025 · similarity 0.87)
- Motion-adaptive Transformer for Event-based Image Deblurring (AAAI 2025 · similarity 0.87)
- EF-3DGS: Event-Aided Free-Trajectory 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.86)
- Talk2Event: Grounded Understanding of Dynamic Scenes from Event Cameras (NeurIPS 2025 · 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.