Research map / Label-efficient and robust learning / Domain adaptation and generalization
EventDance: Unsupervised Source-free Cross-modal Adaptation for Event-based Object Recognition
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 EventDance, an unsupervised source-free image-to-event adaptation framework that reconstructs surrogate images from unlabeled events and transfers source-model knowledge to multi-representation event models.
Problem
Adapting an image-trained classifier to event-camera object recognition when labeled source images are unavailable (only the source model is shared), given the modality gap between images and events.
Approach
A reconstruction-based modality bridging module fine-tunes an event-to-video model (E2VID with EvFlowNet) with entropy and temporal-consistency losses to produce surrogate images for pseudo-labeling via the source model, and a multi-representation knowledge adaptation module trains target models on stack image, voxel grid, and EST representations with cross-representation and cross-modal consisten
Key results
- On N-Caltech101 with ResNet-18, EventDance reaches 66.77% accuracy (+25.96 over the source-only baseline), versus 44.26% for SHOT and 49.33% for SHOT++.
- With ResNet-50 on N-Caltech101 EventDance reaches 92.35%, close to the source-data UDA method DSAN at 94.56%.
- Fine-tuning the reconstruction model increases N-Caltech101 accuracy from 53.04% to 61.83% in the ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EventDance instead of SHOT++ | Image-to-event adaptation · N-MNIST | accuracy: 68.80 -> 71.00; 84.70 -> 86.50; 89.40 -> 92.30 | Helped |
| Use EventDance instead of SHOT++ | Image-to-event adaptation · CIFAR10-DVS | accuracy: 37.26 -> 62.13; 46.37 -> 71.98; 63.41 -> 85.69 | Helped |
| Use EventDance instead of SHOT++ | Image-to-event adaptation · N-CALTECH101 | accuracy: 49.33 -> 66.77; 64.54 -> 72.68; 82.88 -> 92.35 | Helped |
| Add EST as the inference representation instead of voxel grid (ablation) | EventDance · N-CALTECH101 | accuracy: 66.77 -> 66.96; 72.68 -> 73.00; 92.35 -> 92.74 | Helped |
| Use stack image instead of voxel grid at inference (ablation) | EventDance · N-CALTECH101 | accuracy: 66.77 -> 66.70; 72.68 -> 71.16; 92.35 -> 91.54 | Hurt |
| Train target models using all three event representations instead of one representation (ablation) | EventDance · N-CALTECH101 | accuracy: 61.63 -> 66.91; 63.58 -> 66.91; 65.74 -> 66.91 | Helped |
| Fine-tune the reconstruction model FR for surrogate image generation (ablation) | EventDance · N-CALTECH101 | accuracy: 53.04 -> 61.83 | Helped |
Surprising findings
- Fine-tuning the reconstruction model for knowledge extraction improves accuracy even though the reconstructed images look less distinct than before training.
Limitations
- Training three target models with different representations increases training computational cost.
- Building surrogate data is computationally expensive.
- Evaluated only on three small event benchmarks with image-to-event and edge-to-voxel settings.
Experimental setup
Datasets: N-MNIST CIFAR10-DVS N-Caltech101 ImageNet
Benchmarks: N-MNIST CIFAR10-DVS N-Caltech101
Metrics: accuracy
Baselines: SHOT SHOT++ Zhao et al. (CTN) DSAN Ev-LaFOR EV-VGCNN Deep SNN Phased LSTM PointNet++
Builds on: E2VID EvFlowNet SHOT CTN
Closest papers by meaning
- OpenESS: Event-based Semantic Scene Understanding with Open Vocabularies (CVPR 2024 · similarity 0.86)
- PASS: Path-selective State Space Model for Event-based Recognition (NeurIPS 2025 · similarity 0.86)
- EventGPT: Event Stream Understanding with Multimodal Large Language Models (CVPR 2025 · similarity 0.85)
- Cross-modal Multi-task Learning for Multimedia Event Extraction (AAAI 2025 · similarity 0.85)
- Event Stream-based Visual Object Tracking: A High-Resolution Benchmark Dataset and A Novel Baseline (CVPR 2024 · similarity 0.85)
- EventFly: Event Camera Perception from Ground to the Sky (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.