How Far Can Frozen Vision-Language Models Go on Event Data? A Training-Free Study on EventBench
Organizations: Shahjalal University of Science and Technology, Sylhet, Bangladesh · North South University, Dhaka, Bangladesh
Abstract
We present a training-free system for the EBMV @ ECCV 2026 challenge "Event Stream Understanding with EventBench" (General Understanding track), which evaluates vision-language models on five event-camera understanding tasks. We use none of the challenge's training data: our goal is to measure how far off-the-shelf, RGB-pretrained models can go on event data through representation and prompting alone. For four of the five categories (gesture recognition, detailed understanding, causal reasoning, and event OCR) we render event bins as red/ON, blue/OFF frames, select a category-specific subset of bins, stitch them into a single grid image, and query the same frozen Qwen3.5 9B vision-language model with a category-specific prompt for all four tasks. The fifth category, action recognition, feeds every bin directly into the organizers' EventGPT-Plus-2B. The system reaches 62.28% overall accuracy across 2,084 test questions (rank 9/27), led by action recognition and causal reasoning (80.27%, 77.22%) and limited most by gesture recognition and event OCR (54.70%, 34.57%). We visualize test samples from every category together with the exact images the model receives, score six alternative representations against the evaluation server, and analyze the full test set post hoc. That analysis shows that our bounding-box crop almost never tightened around content, so event OCR text reached the model at roughly one fifth of its horizontal resolution, and that detailed-understanding predictions are skewed toward the first two answer options.
Figures & tables
| Method | How events enter the model | Event training |
|---|---|---|
| EventGPT [ 26 ] | event encoder, aggregator, adapter | ✓ three stages |
| EventVL [ 22 ] | event spatio-temporal representation | ✓ 1.4M pairs |
| LET-US [ 11 ] | compressed long-stream features | ✓ two stages |
| EventFlash [ 25 ] | sparsified spatio-temporal tokens | ✓ curriculum |
| Event-MLLM [ 42 ] | event features fused with RGB | ✓ |
| RE-VLM [ 23 ] | parallel RGB and event encoders | ✓ progressive |
| Category | Qs | Answer | Source datasets (per EventBench [ 24 ] ) | Resolution | Bins/clip | Events/bin |
|---|---|---|---|---|---|---|
| Gesture recognition | 894 | 4-way MC | EB-HandGesture [ 1 ] , EHWGesture [ 2 ] | / lower | 8 | 10.2k |
| Detailed understanding | 483 | 4-way MC | Kinetics [ 20 ] , SportsSloMo [ 9 ] , MotionBench [ 18 ] , PLM-Data [ 12 ] , PE-Data [ 4 ] , DSEC [ 17 ] , N-ImageNet [ 21 ] | 16 | 29.1k | |
| Action recognition | 365 | 4-way MC | DailyDVS-200 [ 35 ] , Bullying10K [ 13 ] , UCF-101 [ 33 ] , HARDVS [ 37 ] | , | 16 | 3.5k |
| Causal reasoning a | 180 | 4-way MC | ActivityNet [ 6 ] , Charades [ 32 ] , MotionSight [ 14 ] , WebVid-10M [ 3 ] | 32 | 39.6k | |
| Event OCR | 162 | free text | EventSTR [ 36 ] | 8 | 15.5k b |
| Category | Model | Canvas | k | Crop | Select. |
|---|---|---|---|---|---|
| Gesture | Qwen3.5 9B | 8 | shared | uniform | |
| Detailed | Qwen3.5 9B | 8 | per-bin | uniform | |
| Action | EventGPT-Plus-2B | all | none | none | |
| Causal | Qwen3.5 9B | 6 | shared | uniform | |
| Event OCR | Qwen3.5 9B | 8 | per-bin | count |
| Category | Model | Score |
|---|---|---|
| Gesture recognition | Qwen3.5 9B | 54.70% |
| Detailed understanding | Qwen3.5 9B | 66.46% |
| Action recognition | EventGPT-Plus-2B | 80.27% |
| Causal reasoning | Qwen3.5 9B | 77.22% |
| Event OCR | Qwen3.5 9B | 34.57% |
| Overall (rank 9/27) | 62.28% | |
| Alternative (Qwen3.5 9B) | Category | Score | Adopted | |
|---|---|---|---|---|
| Foveated accumulation [ 5 ] | OCR | 29.63 | 34.57 | 4.94 |
| Accumulation, activity | OCR | 21.60 | 34.57 | 12.97 |
| Accumulation, signed | OCR | 22.22 | 34.57 | 12.35 |
| Category-aware selection | Causal | 55.00 | 77.22 | 22.22 |
| All above threshold | Detailed | 53.83 | 66.46 | 12.63 |
| MMR bin selection | Gesture | 49.47 | 54.70 | 5.23 |
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| Events per bin (percentile) | ||||||
|---|---|---|---|---|---|---|
| Category | File-name groups (clips) | 5 | 25 | 50 | 75 | 95 |
| Gesture | camera* 494, X* 400 | 3.5k | 5.8k | 10.2k | 18.8k | 61.0k |
| Detailed | video-id 300, v_* 45, X* 7, other 131 | 2.0k | 10.5k | 29.1k | 71.1k | 243.6k |
| Action | C#P#* 220, dvSave* 145 | 0.7k | 1.6k | 3.5k | 7.3k | 22.9k |
| Causal | v_* 104 | 2.2k | 12.3k | 39.6k | 96.5k | 779.6k |
| Event OCR | recording* 162 | 0 | 3.5k | 15.5k | 45.1k | 187.0k |
Event-Based Vision
Event-based vision uses asynchronous camera measurements to perceive motion and scene changes with high temporal resolution and dynamic range.
See all 250 papers