Research map / Label-efficient and robust learning / Domain adaptation and generalization
EventFly: Event Camera Perception from Ground to the Sky
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 EventFly, a cross-platform unsupervised domain adaptation framework for event-camera dense semantic segmentation (vehicle, drone, quadruped) using entropy minimization on high-activation regions, activation-guided data blending, and dual-discriminator adversarial alignment, plus the EXPo benchmark.
Problem
Event-camera dense perception degrades when transferred across robot platforms (vehicle, drone, quadruped) that differ in viewpoint, motion dynamics, and class distributions, and conventional frame-based domain adaptation does not fit event data.
Approach
Uses platform-specific event activation regions to (i) minimize prediction entropy in high-activation target regions (EAP), (ii) blend source and target voxel grids via density-similarity masks (EventBlend), and (iii) align source, blended, and target features with two discriminators (EventMatch).
Key results
- EventFly reaches 40.05% mIoU on vehicle-to-quadruped adaptation, a 59.2% relative gain over source-only (25.15%).
- On vehicle-to-drone, EventFly achieves 32.67% mIoU versus 15.04% source-only and 29.69% for the best prior method PLSR.
- Averaged over six cross-platform settings, EventFly reports 71.25% Acc and 36.08% mIoU, best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EventFly instead of source-only training | E2VID event segmentation model · vehicle → drone (Pv → Pd) | mIoU: 15.04 -> 32.67 | Helped |
| Use EventFly instead of source-only training | E2VID event segmentation model · vehicle → quadruped (Pv → Pq) | mIoU: 25.15 -> 40.05 | Helped |
| Use EventFly instead of source-only training | E2VID event segmentation model · drone → vehicle (Pd → Pv) | mIoU: 20.79 -> 39.92 | Helped |
| Use EventFly instead of source-only training | E2VID event segmentation model · drone → quadruped (Pd → Pq) | mIoU: 23.06 -> 37.37 | Helped |
| Use EventFly instead of source-only training | E2VID event segmentation model · quadruped → vehicle (Pq → Pv) | mIoU: 21.30 -> 37.70 | Helped |
| Use EventFly instead of source-only training | E2VID event segmentation model · quadruped → drone (Pq → Pd) | mIoU: 16.85 -> 28.79 | Helped |
| Add EventMatch (ablation) | EventFly, vehicle → drone · vehicle → drone (Pv → Pd) | mIoU: 15.04 -> 25.28 | Helped |
| Add EventBlend (ablation) | EventFly, vehicle → drone · vehicle → drone (Pv → Pd) | mIoU: 15.04 -> 27.91 | Helped |
| Add dual-branch network and EventMatch (ablation) | EventFly, vehicle → drone · vehicle → drone (Pv → Pd) | mIoU: 15.04 -> 30.24 | Helped |
| Add dual-branch network and EventBlend (ablation) | EventFly, vehicle → drone · vehicle → drone (Pv → Pd) | mIoU: 15.04 -> 31.96 | Helped |
| Add EventMatch to dual-branch network and EventBlend (ablation) | EventFly, vehicle → drone · vehicle → drone (Pv → Pd) | mIoU: 31.96 -> 32.67 | Helped |
| Add EventMatch (ablation) | EventFly, vehicle → quadruped · vehicle → quadruped (Pv → Pq) | mIoU: 25.15 -> 33.66 | Helped |
Surprising findings
- The drone platform is the hardest target to adapt to, attributed to its unique motion patterns and perspective-based dynamics.
Limitations
- Baselines are frame-based adaptation methods re-implemented on event data because no event-specific adaptation baselines exist.
- Drone target adaptation remains far below supervised target performance (32.67% vs 42.90% mIoU).
Experimental setup
Datasets: EXPo
Benchmarks: EXPo
Metrics: Acc mAcc mIoU fIoU per-class IoU
Baselines: Source-Only AdaptSegNet CBST IntraDA DACS MIC PLSR
Builds on: E2VID ESS DACS MIC AdaptSegNet
Closest papers by meaning
- EventZoom: A Progressive Approach to Event-Based Data Augmentation for Enhanced Neuromorphic Vision (AAAI 2025 · similarity 0.87)
- Talk2Event: Grounded Understanding of Dynamic Scenes from Event Cameras (NeurIPS 2025 · similarity 0.86)
- Event Stream-based Visual Object Tracking: A High-Resolution Benchmark Dataset and A Novel Baseline (CVPR 2024 · similarity 0.86)
- V2V: Scaling Event-Based Vision through Efficient Video-to-Voxel Simulation (NeurIPS 2025 · similarity 0.86)
- EventDance: Unsupervised Source-free Cross-modal Adaptation for Event-based Object Recognition (CVPR 2024 · similarity 0.85)
- Adapting to the Unknown: Training-Free Audio-Visual Event Perception with Dynamic Thresholds (CVPR 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.