File size: 9,172 Bytes
299f378 a3044f5 299f378 a3044f5 4a993ec a3044f5 299f378 a3044f5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 | ---
# TODO: confirm licence. Annotations are derived from EPIC-KITCHENS / VISOR and HD-EPIC
# (CC BY-NC 4.0) and Ego4D (Ego4D licence agreement). A non-commercial licence is the safe default.
license: cc-by-nc-4.0
language:
- en
pretty_name: BiScopeCap-100K
size_categories:
- 100K<n<1M
task_categories:
- video-text-to-text
- video-classification
tags:
- egocentric
- hand-object-interaction
- sub-action-recognition
- video-captioning
- epic-kitchens
- ego4d
- hd-epic
annotations_creators:
- machine-generated
- expert-generated
source_datasets:
- extended|epic-kitchens-100
- extended|epic-kitchens-visor
- extended|ego4d
- extended|hd-epic
- extended|hoi-ref
configs:
- config_name: default
data_files:
- split: train
path: biscopecap_train.json
- split: validation
path: biscopecap_val.json
# Optional: gate access so users acknowledge the upstream licences.
# extra_gated_prompt: >-
# BiScopeCap contains annotations derived from EPIC-KITCHENS, Ego4D and HD-EPIC.
# Raw videos are NOT included and must be obtained under their original licences.
# extra_gated_fields:
# I agree to use this dataset for non-commercial research only: checkbox
# I have accepted the Ego4D licence agreement: checkbox
---
# BiScopeCap-100K
**BiScopeCap** is a video-level egocentric captioning dataset for **sub-action detection**: recognising the high-level action in a clip *and* the separate roles of the **left** and **right** hands, together with the objects each hand manipulates. Each caption is paired with structured spatial metadata (per-keyframe bounding boxes for hands and objects).
It was introduced in:
> **Sub-actions in Action: Text-Guided Hand-Role Alignment for Sub-Action Recognition**
> Sadegh Rahmaniboldaji, Filip Rybansky, Quoc Vuong, Frank Guerin, Andrew Gilbert
> *BMVC 2026*
> [Paper](TODO) · [Project page](TODO) · [Code](TODO)
<!-- TODO: optional teaser figure, e.g.  -->
## Dataset summary
Most egocentric benchmarks describe only the high-level action ("a person is picking up a pan"). BiScopeCap also captures what each hand does within that action:
> *A person is washing a plate. The left hand holds the plate, while the right hand uses a sponge to clean it.*
Captions were built by standardising and enriching annotations from EPIC-KITCHENS / EPIC-VISOR, Ego4D, HD-EPIC and HOI-QA. They were consolidated into video-level captions with Gemini 2.0 Flash, conditioned on human narrations and annotated hand/object regions. Each caption follows a unified format:
1. the **high-level action**,
2. the **left- and right-hand roles** (when the hand is visible or active),
3. **spatial metadata** in a `<meta>…</meta>` span.
**No videos are distributed in this repository.** Only annotations are released. Clips have to be obtained from the original datasets and cut using the identifiers below (see [Obtaining the videos](#obtaining-the-videos)).
## Dataset structure
### Splits
| Source subset | Train | Test |
|---|---:|---:|
| EPIC-KITCHENS / VISOR | 31,432 | 3,729 |
| Ego4D | 22,916 | 3,304 |
| HD-EPIC | 38,520 | 2,846 |
| **Total** | **92,868** | **9,879** |
<!-- TODO: the file names say "val" but the paper calls this split "test". Keep whichever name matches the paper. -->
### Data fields
Each record follows the LLaVA / VideoChatGPT conversation format:
```json
{
"id": "P01_107_0",
"video": "P01_107_0",
"key_frames": [21, 58],
"conversations": [
{"from": "human", "value": "describe the action in the video. \n<video>"},
{"from": "gpt", "value": "A person is closing a door. The right hand is holding and manipulating the door to close it. <meta>[{\"frame\": 21, \"object\": \"door\", \"bbox\": [118, 0, 106, 199]}, {\"frame\": 21, \"object\": \"right hand\", \"bbox\": [155, 151, 21, 73]}, ...]</meta>"}
]
}
```
| Field | Type | Description |
|---|---|---|
| `id` | string | Unique clip identifier. Its format shows the source subset (see below). |
| `video` | string | Clip file name (without extension) that the record refers to. |
| `key_frames` | list[int] | Frame indices in the clip that the spatial metadata refers to. |
| `conversations` | list[dict] | One human prompt plus one `gpt` target caption. |
**Caption and metadata.** The target caption is natural language followed by a `<meta>…</meta>` span holding a JSON list of entries:
| Key | Description |
|---|---|
| `frame` | Keyframe index (one of `key_frames`). |
| `object` | Object label, including `"left hand"` / `"right hand"`. |
| `bbox` | `[x, y, w, h]` in pixels, in the **224 × 224** resized frame. |
To get the plain caption, strip everything from `<meta>` onwards. Every record has a well-formed `<meta>` span.
**ID formats by source**
| Source | Example `id` |
|---|---|
| EPIC-KITCHENS / VISOR | `P01_107_0` (`<video_id>_<narration_index>`) |
| HD-EPIC | `P01-20240202-110250-1` |
| Ego4D | `<video_uid>_<clip_uid>_<annotation_uid>` |
<!-- TODO: confirm the exact meaning of each ID component and of the key_frames reference (clip-relative vs. source-video frame). -->
### Loading
```python
from datasets import load_dataset
import json, re
ds = load_dataset("TODO-org/BiScopeCap") # TODO: repo id
def split_caption(text):
m = re.search(r"<meta>(.*)</meta>", text, re.S)
caption = text[: m.start()].strip() if m else text.strip()
meta = json.loads(m.group(1)) if m else []
return caption, meta
caption, meta = split_caption(ds["train"][0]["conversations"][1]["value"])
```
## Obtaining the videos
BiScopeCap does not redistribute any video. To use it:
1. Download the source videos under their own licences:
- EPIC-KITCHENS-100 / VISOR: https://epic-kitchens.github.io
- Ego4D: https://ego4d-data.org (licence agreement required)
- HD-EPIC: https://hd-epic.github.io
2. Trim clips at the narration timestamps and resize them to the model input resolution (224 × 224 in our experiments).
3. Name each clip after its `video` field.
<!-- TODO: link the clip-extraction script from the code repository, or add a per-clip timestamp file (start/end time per id) to this repo. -->
## Dataset creation
### Curation pipeline
1. **Per-source processing.** EPIC-KITCHENS and Ego4D supply narrations and temporal boundaries. EPIC-VISOR segmentations, processed by HOI-QA, supply hand/object boxes and keyframe captions. HD-EPIC supplies video-level hand-motion descriptions, reformulated to add the missing high-level action and structured metadata.
2. **Caption generation.** Gemini 2.0 Flash (May–Sep 2025) summarises these signals into one video-level caption per clip. It is conditioned on the human narration and the annotated regions, so the hand-role attribution stays grounded.
3. **Video cropper and caption adjuster.** Clips are segmented by narration timestamps and resized. Bounding boxes are re-aligned spatially and temporally to the resized keyframes.
### Statistics
- 12,814 action instances and 5,727 active objects, including hands.
- Most frequent objects other than hands: *pan* (4,294), *tap* (4,271), *knife* (4,271), *sink* (3,180), *plate* (3,073).
- Left-hand, right-hand and bimanual interactions are all well represented. A small number of captions mention no hand.
<!-- TODO: optionally add the word cloud and the hand-distribution figure from the supplementary material. -->
### Annotation quality
A manual audit of 200 randomly sampled training clips, balanced across sources, found:
- ≈5% with minor descriptive inaccuracies (e.g. imprecise verbs),
- ≈8% with left-hand hallucination,
- ≈10% with right-hand hallucination.
Captions are machine-generated and should be treated as **noisy supervision**. They are not exhaustive ground truth.
## Uses
**Intended uses**
- Training and evaluating video-language models for fine-grained, hand-aware egocentric captioning.
- Sub-action detection: identifying the left- and right-hand roles within a high-level action.
- Research on grounded captioning that uses the spatial `<meta>` annotations.
**Out-of-scope uses**
- Commercial use, subject to the upstream licences.
- Identifying or profiling individuals in the source recordings.
- Using the captions as verified ground truth in safety-critical settings. They are machine-generated and contain noise.
## Licence
The annotations in this repository are released under **TODO (e.g. CC BY-NC 4.0)**. Using them also requires following the licences of the source datasets: EPIC-KITCHENS / VISOR, Ego4D and HD-EPIC. No raw video is redistributed.
## Citation
```bibtex
@inproceedings{rahmaniboldaji2026subactions,
title = {Sub-actions in Action: Text-Guided Hand-Role Alignment for Sub-Action Recognition},
author = {Rahmaniboldaji, Sadegh and Rybansky, Filip and Vuong, Quoc and Guerin, Frank and Gilbert, Andrew},
booktitle = {British Machine Vision Conference (BMVC)},
year = {2026}
}
```
Please also cite the source datasets: EPIC-KITCHENS-100, EPIC-KITCHENS VISOR, Ego4D, HD-EPIC and HOI-Ref.
## Acknowledgements
This work was funded by the Leverhulme Trust Research Project Grant RPG-2023-079.
|