Kavin60606/egocentric-affordance-dataset

Egocentric Affordance Dataset

A large-scale dataset of 48,400 annotated egocentric video frames across 5 real-world activity categories, annotated with fine-grained affordance labels using Qwen3-VL-2B vision-language model. Designed for training and evaluating embodied AI systems, robotic manipulation models, and vision-language models (VLMs) that reason about human-object interactions.

---

Dataset Summary

Metric	Value
Total Frames	48,400
Valid Affordance Labels	45,682 (94.4%)
Unique Affordance Labels	5,025
Categories	5
Source Videos	250 YouTube videos
View	Egocentric (first-person)
Annotation Model	Qwen3-VL-2B (checkpoint-3500)
License	CC-BY 4.0

---

Categories

Batch	Category	Videos	Frames	Valid Affordances
1	cleaning_household	50	8,600	8,136
2	workshop_diy	50	9,800	9,449
3	factory_working	50	10,000	9,475
4	gardening_outdoor	50	10,000	9,182
5	medical_healthcare	50	10,000	9,440
Total		250	48,400	45,682

---

Affordance Label Format

Each annotation uses the format <verb> <object>, where the verb describes the manipulation action and the object describes the target item.

Example labels:

grasp handle
press button
open door
lift box
insert needle
cut grass
pour liquid
rotate knob
slide drawer
remove cap

Top 20 most frequent affordances:

Affordance	Count
lift hand	1,101
press button	1,094
open door	960
grasp hand	790
grasp tool	662
grasp object	626
grasp hands	624
open mouth	536
grasp handle	466
insert needle	438
lift box	374
grasp steering wheel	373
slide car	309
grasp cup handle	286
grasp phone	285
lift plant	283
lift head	279
grasp rope	272
cut grass	258
grasp plant	242

---

Data Structure

affordance_batches_1_to_5.zip
└── batches/
    ├── batch_1/  (cleaning_household)
    │   ├── annotations.jsonl      # Frame-level affordance annotations
    │   ├── search_log.json        # Video metadata
    │   ├── manifest.csv           # Frame manifest
    │   └── frames/
    │       └── <video_id>/
    │           └── <video_id>_XXXX.jpg   # Extracted frames (2.4s intervals)
    ├── batch_2/  (workshop_diy)
    ├── batch_3/  (factory_working)
    ├── batch_4/  (gardening_outdoor)
    └── batch_5/  (medical_healthcare)

JSONL Annotation Schema

Each line in annotations.jsonl is a JSON object:

{
  "video_id": "8fVF6iDnVTQ",
  "frame_id": 1,
  "timestamp": 2.41,
  "image_path": "frames/8fVF6iDnVTQ/8fVF6iDnVTQ_0001.jpg",
  "affordance": "remove plate",
  "raw_caption": "",
  "task": "cleaning",
  "view": "egocentric",
  "source": "youtube",
  "model": "qwen3-vl-2b-instruct-4bit",
  "prompt_version": "affordance_v2_expanded_verbs"
}

Field descriptions:

Field	Type	Description
video_id	string	YouTube video ID
frame_id	int	Frame index within video (1-based)
timestamp	float	Frame timestamp in seconds
image_path	string	Relative path to frame image inside zip
affordance	string	Affordance label (<verb> <object>) or "none"
task	string	High-level task category (cleaning, diy, factory, gardening, medical)
view	string	Always "egocentric"
source	string	Always "youtube"
model	string	Annotation model used
prompt_version	string	Prompt template version

---

Usage

Load annotations

import json
from pathlib import Path

annotations = []
for batch_num in range(1, 6):
    jsonl_path = f"batches/batch_{batch_num}/annotations.jsonl"
    with open(jsonl_path) as f:
        for line in f:
            annotations.append(json.loads(line))

print(f"Total annotations: {len(annotations)}")
# Filter valid affordances
valid = [a for a in annotations if a["affordance"] != "none" and a["affordance"]]
print(f"Valid affordances: {len(valid)}")

Load with HuggingFace datasets

from datasets import load_dataset

# Load annotations as a dataset
ds = load_dataset("Kavin60606/egocentric-affordance-dataset", data_files="batches/batch_1/annotations.jsonl")
print(ds)

Extract verb/object pairs

import json

verbs = set()
objects = set()

with open("batches/batch_1/annotations.jsonl") as f:
    for line in f:
        rec = json.loads(line)
        aff = rec.get("affordance", "")
        if aff and aff != "none":
            parts = aff.split(" ", 1)
            verbs.add(parts[0])
            if len(parts) > 1:
                objects.add(parts[1])

print(f"Unique verbs: {sorted(verbs)}")

Filter by category

import json

category_frames = {}
for batch_num in range(1, 6):
    path = f"batches/batch_{batch_num}/annotations.jsonl"
    with open(path) as f:
        frames = [json.loads(line) for line in f]
    task = frames[0]["task"] if frames else f"batch_{batch_num}"
    category_frames[task] = frames
    print(f"{task}: {len(frames)} frames")

---

Annotation Methodology

Frames were extracted from egocentric YouTube videos at ~2.4 second intervals using OpenCV. Each frame was annotated using Qwen3-VL-2B (4-bit quantized, fine-tuned checkpoint at step 3500) with the prompt:

"Describe the single most salient affordance visible from the egocentric perspective. Use the format: . Valid verbs include: grasp, pull, push, rotate, press, lift, slide, place, open, close, insert, remove, pour, cut, stir, scoop, squeeze, turn_on, turn_off."

Annotations with "none" indicate frames where no clear affordance was detectable (e.g., transitional frames, occlusion).

---

Source Videos

Videos were sourced from YouTube using keyword searches targeting egocentric/first-person/POV content within each category. Video IDs are recorded in search_log.json per batch. All videos are publicly available on YouTube.

Note: This dataset contains extracted frames and annotations only. The original videos are not redistributed. Users are responsible for complying with YouTube's Terms of Service for any use of video content.

---

Intended Uses

• Robotic manipulation: Training affordance-aware manipulation policies
• Embodied AI: Grounding language instructions to visual affordances
• VLM fine-tuning: Supervised fine-tuning for egocentric understanding
• Action anticipation: Predicting next actions from first-person video
• Human-robot interaction: Understanding human manipulation for robot imitation

---

Limitations

• Annotations are model-generated (Qwen3-VL-2B) and may contain errors (~5.6% invalid/none)
• Video quality varies (720p–1080p); some frames may be blurry or poorly lit
• medical_healthcare batch task labels may show "unknown" in some records
• Coverage is biased toward English-language YouTube content

---

Citation

If you use this dataset, please cite:

@dataset{kavinraj2026egocentric,
  author    = {Kavin Raj},
  title     = {Egocentric Affordance Dataset},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/Kavin60606/egocentric-affordance-dataset}
}

---

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC-BY 4.0) license.

Created by Kavin Raj as part of the Fidelity Dynamics Multimodal Data Foundry project.