Simplify dataset README
Browse files
README.md
CHANGED
|
@@ -1,15 +1,4 @@
|
|
| 1 |
-
-
|
| 2 |
-
pretty_name: GuideMe-Test
|
| 3 |
-
language:
|
| 4 |
-
- en
|
| 5 |
-
tags:
|
| 6 |
-
- video
|
| 7 |
-
- benchmark
|
| 8 |
-
---
|
| 9 |
-
|
| 10 |
-
# GuideMe: Multi-Domain Task Guidance and Intervention in Streaming Video
|
| 11 |
-
|
| 12 |
-
Accepted at **ECCV 2026**.
|
| 13 |
|
| 14 |
GuideMe is a benchmark for training and evaluating multimodal large language models as real-time procedural coaches. It covers next-step instruction, completion feedback, error detection, and corrective guidance across cooking, daily-life tasks, object manipulation, and fitness.
|
| 15 |
|
|
@@ -17,26 +6,6 @@ GuideMe is a benchmark for training and evaluating multimodal large language mod
|
|
| 17 |
- Code: https://github.com/fawnliu/GuideMe
|
| 18 |
- Paper: https://arxiv.org/abs/2607.02991
|
| 19 |
|
| 20 |
-
## This release
|
| 21 |
-
|
| 22 |
-
This repository contains the **GuideMe-Test** split: **473 videos**, **46.7 hours**.
|
| 23 |
-
|
| 24 |
-
| Domain | Source | Videos |
|
| 25 |
-
| --- | --- | --- |
|
| 26 |
-
| Cooking | captaincook4d | 92 |
|
| 27 |
-
| Daily life | egoper | 100 |
|
| 28 |
-
| Object manipulation | holoassist | 207 |
|
| 29 |
-
| Fitness | fitness | 74 |
|
| 30 |
-
|
| 31 |
-
Files:
|
| 32 |
-
|
| 33 |
-
- `annotations.jsonl`: one JSON object per line. Each record has `video_id`, `conversations` (`from`, `value`, `timestamp`), and `videos`. Some records also have `activity_name`.
|
| 34 |
-
- `videos_000.tar` … `videos_003.tar`: the mp4 files. Each archive is under 9.5 GB. Extracting them gives a flat directory of videos; the filename is the basename of the path in `videos`.
|
| 35 |
-
|
| 36 |
-
The full benchmark has 2,458 videos (1,985 train / 473 test), 223.7 hours, and 47,775 streaming interaction samples. Train is not included here.
|
| 37 |
-
|
| 38 |
-
## Citation
|
| 39 |
-
|
| 40 |
```bibtex
|
| 41 |
@article{liu2026guideme,
|
| 42 |
title={GuideMe: Multi-Domain Task Guidance and Intervention in Streaming Video},
|
|
|
|
| 1 |
+
# GuideMe: Multi-Domain Task Guidance and Intervention in Streaming Video (ECCV 2026)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
|
| 3 |
GuideMe is a benchmark for training and evaluating multimodal large language models as real-time procedural coaches. It covers next-step instruction, completion feedback, error detection, and corrective guidance across cooking, daily-life tasks, object manipulation, and fitness.
|
| 4 |
|
|
|
|
| 6 |
- Code: https://github.com/fawnliu/GuideMe
|
| 7 |
- Paper: https://arxiv.org/abs/2607.02991
|
| 8 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
```bibtex
|
| 10 |
@article{liu2026guideme,
|
| 11 |
title={GuideMe: Multi-Domain Task Guidance and Intervention in Streaming Video},
|