This dataset contains ego+exo video pairs and standalone sessions from real shoe manufacturing operations. Please fill in the form below to request access.

Ego-Exo Manufacturing Dataset

Ego-Exo Manufacturing is a video dataset of real shoe manufacturing operations with two configs.

synced_pairs — temporally synchronized ego+exo pairs with expert annotations. Videos are in a separate gated repo: skill-ai/ego-exo-manufacturing-synced.
standalone contains individual sessions from workers, with action labels.

Dataset Statistics

Attribute	Value
Total video content	~275 hours
Synced ego-exo pairs	40 groups
Synced ego duration	~36 hours
Synced exo duration	~36 hours
Standalone ego sessions	80 (~101 hours)
Standalone exo sessions	137 (~102 hours)
Storage size	~1.1 TiB
Ego frame rate	30 fps
Exo frame rate	~25–30 fps (varies per group)
Resolution	1080p
Format	H.264 / MP4
Audio	No
Face privacy	Exo faces Gaussian-blurred
Annotations	Atomic action descriptions, expert commentary

Dataset Structure

The dataset has two configs: synced_pairs (temporally aligned ego+exo pairs with full annotations) and standalone (individual unmatched sessions with action labels).

ego-exo-manufacturing/
├── synced_pairs/
│   ├── takes.json                          # Root metadata — one entry per group
│   ├── takes/
│   │   └── groupXX/
│   │       ├── ego01.mp4                   # Trimmed first-person video
│   │       └── exo01_blurred.mp4           # Trimmed overhead video (faces blurred)
│   └── annotations/
│       ├── atomic_descriptions/groupXX.json
│       └── expert_commentary/groupXX.json
└── standalone/
    ├── ego/NNN/*.mp4                       # NNN = 001–080
    ├── exo/NNN/*.mp4                       # NNN = 001–137
    └── labels/
        ├── ego/NNN.json
        └── exo/NNN.json

`takes.json`

Root-level index for synced_pairs. Both videos in each group are trimmed so that frame 0 of ego01.mp4 and frame 0 of exo01_blurred.mp4 correspond to the same real-world moment.

[
  {
    "group_id": "group34",
    "ego_video": "takes/group34/ego01.mp4",
    "exo_video": "takes/group34/exo01_blurred.mp4",
    "exo_blurred": true,
    "ego_duration_s": 3670.018,
    "ego_fps": 30.0,
    "exo_fps": 26.336
  }
]

`annotations/atomic_descriptions/groupXX.json`

Timestamped action narrations in Ego-Exo4D atomic description format, transcribed from worker narration audio.

{
  "annotations": {
    "group01_ego_shoe_001": [
      {
        "annotation_uid": "9a9d8b89-1a7e-580a-8434-fc13586af3c0",
        "annotator_id": "worker_annotator_001",
        "rejected": false,
        "reject_reason": null,
        "descriptions": [
          {
            "text": "C folds the foam paper with both hands.",
            "timestamp_sec": 2.0,
            "narration_subject": "C",
            "ego_visible": true,
            "unsure": false,
            "_ext": {
              "hand_used": "both",
              "is_essential": true,
              "notes": "An older worker nearby was teaching, so this took longer."
            }
          }
        ]
      }
    ]
  }
}

`annotations/expert_commentary/groupXX.json`

Rich structured expert annotation covering task description, procedural keysteps, proficiency assessment, and mistake labels.

{
  "meta": {
    "dataset": "Shoe Manufacturing v1",
    "annotation_method": "worker_narration_transcribed"
  },
  "take_uid": "group01",
  "task_name": "Shoe box packaging paper folding and placement",
  "scenario_name": "shoe_packaging",
  "action_descriptions": {
    "annotator_id": "worker_annotator_001",
    "descriptions": [
      { "text": "The worker folds the packaging paper in half...", "type": "task_overview" },
      { "text": "Align the short sides facing each other, then fold.", "type": "how" },
      { "text": "The purpose of folding is to create a barrier...", "type": "why" }
    ]
  },
  "proficiency_scores": {
    "overall": 4.3,
    "dimensions": {
      "speed":     { "score": 5, "max": 5 },
      "precision": { "score": 4, "max": 5 },
      "fluency":   { "score": 4, "max": 5 }
    }
  },
  "skill_level": {
    "label": "proficient",
    "reasoning": "Worker was being taught at the beginning but quickly became proficient."
  },
  "mistake_labels": {
    "has_mistakes": false,
    "annotations": [],
    "common_failure_modes": [
      { "description": "Not aligning edges before folding.", "severity": "minor" }
    ]
  },
  "procedure": {
    "keysteps": [
      { "step_id": 1, "label": "Align short edges facing each other" },
      { "step_id": 2, "label": "Fold along short edge" }
    ],
    "dependencies": [
      { "from": 1, "to": 2, "type": "required" }
    ],
    "observed_order": [1, 2, 3, 4, 5, 6],
    "deviations_from_standard": []
  },
  "expert_commentary": {
    "commentary_data": [
      { "text": "The most common mistake is not aligning edges before folding.", "type": "tip_for_improvement" },
      { "text": "Overall: the worker is fast and proficient.", "type": "overall_assessment" }
    ]
  }
}

`standalone/labels/ego|exo/NNN.json`

Per-session action label.

{
  "stage_name": "cementing sole",
  "shoe_component": "sole",
  "activity": "applying adhesive",
  "tools_or_materials": ["cement brush", "rubber sole", "adhesive"],
  "camera_perspective": "ego",
  "hands_visible": true,
  "anomalies": ["none"],
  "confidence": "high",
  "sop": "Worker applies adhesive evenly to the shoe sole contact surface before pressing."
}

Loading the Dataset

Load metadata

import json
from huggingface_hub import hf_hub_download

takes_path = hf_hub_download(
    repo_id="skill-ai/ego-exo-manufacturing",
    filename="synced_pairs/takes.json",
    repo_type="dataset",
)
with open(takes_path) as f:
    takes = json.load(f)

print(f"{len(takes)} groups")
# 40 groups

Load a synced video pair

import cv2
from huggingface_hub import hf_hub_download

REPO = "skill-ai/ego-exo-manufacturing"
GROUP = "group34"

ego_path = hf_hub_download(REPO, f"synced_pairs/takes/{GROUP}/ego01.mp4", repo_type="dataset")
exo_path = hf_hub_download(REPO, f"synced_pairs/takes/{GROUP}/exo01_blurred.mp4", repo_type="dataset")

ego_cap = cv2.VideoCapture(ego_path)
exo_cap = cv2.VideoCapture(exo_path)

# Frame N from ego_cap == Frame N from exo_cap (same real-world moment)
ret_e, ego_frame = ego_cap.read()
ret_x, exo_frame = exo_cap.read()

Load annotations

import json
from huggingface_hub import hf_hub_download

REPO = "skill-ai/ego-exo-manufacturing"
GROUP = "group34"

# Atomic descriptions — timestamped narrations
ann_path = hf_hub_download(REPO, f"synced_pairs/annotations/atomic_descriptions/{GROUP}.json", repo_type="dataset")
with open(ann_path) as f:
    ann = json.load(f)

for key, entries in ann["annotations"].items():
    for entry in entries:
        for desc in entry["descriptions"]:
            print(f"t={desc['timestamp_sec']:.1f}s  {desc['text']}")

# Expert commentary — proficiency, keysteps, mistakes
ec_path = hf_hub_download(REPO, f"synced_pairs/annotations/expert_commentary/{GROUP}.json", repo_type="dataset")
with open(ec_path) as f:
    ec = json.load(f)

print(f"Task: {ec['task_name']}")
print(f"Skill level: {ec['skill_level']['label']}  ({ec['proficiency_scores']['overall']}/5)")
for step in ec["procedure"]["keysteps"]:
    print(f"  Step {step['step_id']}: {step['label']}")

License

Licensed under Apache 2.0.

Citation

@dataset{egosexomanufacturing2026,
  title  = {Ego-Exo Manufacturing},
  year   = {2026},
  url    = {https://huggingface.co/datasets/skill-ai/ego-exo-manufacturing},
}

Downloads last month: 65

Collection including skill-ai/ego-exo-manufacturing

Ego-Exo Manufacturing

Collection

Ego+exo video of real manufacturing: synchronized video pairs, atomic action labels, skill assessment. For robotics and activity recognition research. • 2 items • Updated 2 days ago