README.md

---
pretty_name: EgoIntent
language:
  - en
task_categories:
  - video-text-to-text
size_categories:
  - 1K<n<10K
source_datasets:
  - extended
tags:
  - video
  - egocentric-video
  - intent-understanding
  - procedural-reasoning
  - action-anticipation
  - multimodal
configs:
  - config_name: default
    drop_labels: true
---

# EgoIntent

## Dataset Summary

EgoIntent is an open-ended benchmark for understanding intent at pre-outcome micro-steps in egocentric procedural activities. Given only the visual evidence available before a step's key outcome, a model must infer:

1. **Local Intent (What):** the actor's immediate goal in the current micro-step.
2. **Procedural Intent (Why):** how the current micro-step advances the broader procedure.
3. **Next-Plan (Next):** the action most likely to occur immediately afterward.

The benchmark contains **3,014 manually annotated micro-step clips** constructed from **32 Ego4D source videos/event directories** across **15 indoor and outdoor daily-life scenarios**. The activities include cooking, cleaning, organizing, repairing, painting, gardening, and outdoor manual work.

Each released MP4 ends at a manually selected pre-outcome cutoff. Outcome-revealing frames and subsequent actions are excluded from the model input, while the immediate observed continuation is retained only as a text target.

## Benchmark at a Glance

| Item | Count |
| --- | ---: |
| Pre-outcome MP4 clips / micro-steps | 3,014 |
| JSON annotation files | 32 |
| Ego4D source videos / event directories | 32 |
| Scenarios | 15 |
| Indoor scenarios | 10 |
| Outdoor scenarios | 5 |
| Total repository data | approximately 17 GB |

There are no predefined train, validation, or test splits in this release.

## Task Definition

For each micro-step, the input is the egocentric observation from `start_time` to `obs_end_time`. A model generates three open-ended text predictions:

- **Local Intent:** purpose at the current micro-step level, rather than a literal paraphrase of visible motion.
- **Procedural Intent:** the functional role of the step in the surrounding activity.
- **Next-Plan:** the immediate action expected after the observed clip.

The three targets are related but non-redundant. Because valid descriptions may differ in wording, and multiple futures can be reasonable, evaluation should use semantic agreement rather than exact string matching.

## Directory Structure

```text
EgoIntent/
├── metadata.jsonl
├── indoor/
│   └── art_studio/
│       └── draw/
│           ├── step1.mp4
│           ├── step2.mp4
│           ├── ...
│           └── step_label.json
└── outdoor/
    └── garden/
        └── cut_trees/
            ├── step1.mp4
            ├── step2.mp4
            ├── ...
            └── step_label.json
```

Paths follow:

```text
{setting}/{scene}/{event}/step{step_id}.mp4
{setting}/{scene}/{event}/step_label.json
```

where `setting` is `indoor` or `outdoor`.

### Dataset Viewer Metadata

The root-level `metadata.jsonl` contains one row per MP4 and links the video through its relative `file_name`. Hugging Face renders that field as the `video` column. The remaining Viewer columns are:

- `id`: a globally unique ID in the form `{setting}/{scene}/{event}/step{step_id}`;
- `local_intent`;
- `procedural_intent`;
- `observed_next_step`;
- `plausible_next_steps`.

The Dataset Card configuration explicitly sets `drop_labels: true`, so event directory names are not interpreted as classification labels.

## Annotation Format

Each event directory contains one `step_label.json` file:

```json
{
  "video_uid": "8bef8b33-a1ee-43e7-a2d0-8cf68653cb0c",
  "scene": "art_studio",
  "event": "draw",
  "steps": [
    {
      "step_id": 1,
      "start_time": 0,
      "obs_end_time": 0.12332,
      "local_intent": "hold the sketch paper",
      "procedural_intent": "align reference for painting",
      "observed_next_step": "release the paper on the table",
      "plausible_next_steps": [
        "lift the sketch paper",
        "view the phone screen"
      ]
    }
  ]
}
```

### Field Definitions

| Field | Type | Description |
| --- | --- | --- |
| `video_uid` | string | Identifier of the Ego4D source video. |
| `scene` | string | Daily-life scenario associated with the source video. |
| `event` | string | Procedural activity/event directory name. |
| `steps` | list | Micro-step annotations for this source video/event. |
| `step_id` | integer | Step identifier used by the matching `step{step_id}.mp4` file. |
| `start_time` | number | Start timestamp of the observation in the source video, in seconds. |
| `obs_end_time` | number | Manually selected pre-outcome cutoff timestamp, in seconds. |
| `local_intent` | string | Immediate goal pursued in the current micro-step. |
| `procedural_intent` | string | Role of the micro-step in the broader procedure. |
| `observed_next_step` | string | Immediate continuation observed after the cutoff; used as the canonical Next-Plan target. |
| `plausible_next_steps` | list of strings | Other reasonable continuations supported by the pre-outcome evidence. |

## Scenario Composition

| Setting | Scenario | Event directories | Micro-steps |
| --- | --- | --- | ---: |
| Indoor | `art_studio` | `draw` | 90 |
| Indoor | `bedroom` | `clean_windows`, `sort_clothes` | 191 |
| Indoor | `clothing_closet` | `iron_clothes` | 104 |
| Indoor | `garage` | `assemble_wheels`, `repair_bicycle`, `repair_car` | 321 |
| Indoor | `hallway` | `paint_wall` | 101 |
| Indoor | `kitchen` | `bake_pastry`, `boil_noodles`, `knead_dough`, `make_cakes`, `wash_dishes` | 396 |
| Indoor | `laundry_room` | `wash_clothes`, `wash_clothes_2` | 106 |
| Indoor | `living_room` | `eat`, `organize_dishes` | 198 |
| Indoor | `study_room` | `sort_books`, `tidy_up` | 202 |
| Indoor | `workshop` | `assemble_mower`, `assemble_wood`, `examine_engine`, `layout`, `make_crafts` | 459 |
| Outdoor | `deck` | `measure` | 106 |
| Outdoor | `farm` | `farm_work`, `pick_fruits` | 214 |
| Outdoor | `garden` | `cut_branches`, `cut_trees`, `paint_timber` | 309 |
| Outdoor | `space` | `draw_flowers` | 113 |
| Outdoor | `yard` | `cut_sticks` | 104 |
| **Total** | **15 scenarios** | **32 event directories** | **3,014** |

## Loading the Data

Download the complete repository:

```python
from huggingface_hub import snapshot_download

dataset_root = snapshot_download(
    repo_id="py2279105943/EgoIntent",
    repo_type="dataset",
)
```

To download only one event:

```python
from huggingface_hub import snapshot_download

event_root = snapshot_download(
    repo_id="py2279105943/EgoIntent",
    repo_type="dataset",
    allow_patterns=["indoor/art_studio/draw/*"],
)
```

Read annotations and resolve the matching clip:

```python
import json
from pathlib import Path

event_dir = Path(dataset_root) / "indoor" / "art_studio" / "draw"
annotations = json.loads((event_dir / "step_label.json").read_text())

for step in annotations["steps"]:
    video_path = event_dir / f"step{step['step_id']}.mp4"
    print(video_path, step["local_intent"])
```

This repository uses a hierarchical benchmark layout rather than a packaged `datasets` builder. Loading with `snapshot_download` preserves the paths needed to pair every MP4 with its event-level JSON annotations.

## Construction and Quality Control

Source videos were selected for coherent procedural transitions and clear hand-object interactions. Annotators manually segmented each source video into micro-steps, identified the outcome completing each immediate goal, and placed the observation boundary immediately before that outcome became visually explicit.

Intent labels and temporal boundaries underwent multiple rounds of human review. Reviewers checked that each micro-step expressed one coherent immediate goal, stayed within the intended temporal limit, ended before the key outcome, and maintained distinct abstraction levels across Local Intent, Procedural Intent, and Next-Plan.

## Intended Uses

EgoIntent is intended for research on:

- egocentric video-language understanding;
- step-level intent and goal inference;
- procedural reasoning;
- action anticipation;
- temporal-context and boundary-cue analysis;
- evaluation of multimodal large language models.

The dataset is not intended for identity recognition, surveillance, or inferring sensitive personal attributes.

## Limitations

- The benchmark contains 32 source videos, so multiple micro-steps share the same source context.
- No official train/validation/test split is provided. If constructing splits, separate examples by `video_uid`—and preferably by scene or event—to reduce source-video leakage.
- Labels are written in English and may reflect annotator wording preferences.
- Next actions can be inherently ambiguous. `plausible_next_steps` captures some alternatives but is not exhaustive.
- The benchmark focuses on procedural daily-life activities and should not be treated as representative of all human intentions, environments, or cultures.
- The videos originate from Ego4D; users should consider the privacy, geographic, demographic, and collection biases documented for the source dataset.

## License and Usage Terms

No standalone license is declared for this repository at the time of this release. The underlying videos originate from Ego4D and remain subject to the applicable Ego4D license and usage terms. Users should verify those terms and confirm annotation reuse conditions with the dataset maintainers before redistribution or commercial use.

## Citation

Citation information will be added after the associated paper is publicly available.

## Contact

Questions, corrections, and annotation issues can be reported through the Hugging Face repository's Community tab.