⚡ TL;DR: SPOT-Bench requires a streaming model to monitor a live video stream and proactively decide when to respond. Every prediction across the full video is evaluated using the Timeliness-F1 metric.

Detection tasks (ABD, PNR) are released in full. For Interaction (SQA, SPG) and Intervention (SI, UI) we release a representative validation set while holding out the test set for an upcoming challenge. Stay tuned for updates on our webpage.

spot-bench/
  abd.json          # Action Boundary Detection
  pnr.json          # Point-of-No-Return Detection
  sqa.json          # Streaming Question Answering
  spg.json          # Streaming Procedural Guidance
  si.json           # Solicited Intervention
  ui.json           # Unsolicited Intervention
  videos.zip        # 662 MP4 videos, ~40 GB
  README.md

Benchmark Statistics

Category	Task	File	Videos	Turns	Slots
Detection (full)	ABD	`abd.json`	322	1,621	2,340
Detection (full)	PNR	`pnr.json`	286	1,159	1,234
Interaction (val)	SQA	`sqa.json`	15	95	101
Interaction (val)	SPG	`spg.json`	13	13	95
Intervention (val)	SI	`si.json`	13	13	40
Intervention (val)	UI	`ui.json`	13	13	35
Total			662	2,914	3,845

Notes

Detection tasks are closed-vocabulary. The expected response is specified in the question and is a single token, enabling fast, deterministic evaluation with no LLM-judge required.
Video-only. All audio tracks have been removed. SPOT-Bench focuses exclusively on streaming video understanding and visual proactivity.
Original IDs. Filenames retain the original source-dataset video IDs, allowing each entry to be traced back to its source.

Evaluation

Evaluation code, baselines, and scoring metrics are provided on Github. Place the unzipped videos/ directory and the six JSON files under data/

License

SPOT-Bench is released under the CC BY-NC-SA 4.0 license. For video sources, please refer to the original dataset licenses: Ego4D, Ego-Exo4D, HTStep, HoloAssist, EgoBlind, MovieNet, Perception Test and THUMOS14.

Citation

@article{chatterjee2026don,
  title={Don't Pause! Every prediction matters in a streaming video},
  author={Chatterjee, Dibyadip and Pang, Zhanzhong and Sener, Fadime and Song, Yale and Yao, Angela},
  journal={arXiv preprint arXiv:2604.24317},
  year={2026}
}

Downloads last month: 23

Total file size:

43.2 GB

Paper for cvml-nus/spot-bench

Don't Pause! Every prediction matters in a streaming video

Paper • 2604.24317 • Published Apr 27

You need to agree to share your contact information to access this dataset

SPOT-Bench

Contents