iMED: A Multi-Endoscope Dataset for Surgical 3D Perception

iMED provides recordings from two stereo endoscopes for research in surgical 3D perception (tasks like pose estimation, depth estimation, feature matching, and rigid and deformable novel view synthesis).

Paper (ECCV 2026) · Browse and download files · Dataset DOI · Code and dataloader

Dataset

The full v2.0 release contains 341 sequences across 43 scenes and seven sessions (~48 GB), with ex-vivo, in-vivo post-mortem, and live porcine sequences.

Tissue type	Sessions	Folder pattern
Ex-vivo	001–003	session_00[123]
In-vivo post-mortem	004–006	session_00[456]
Live porcine	007	session_007

• Processed and raw stereo videos from two endoscopes, so each timepoint has 4 views.
• Capture timestamps, camera intrinsics, and calibration files.
• Camera poses, instrument masks, and sequence metadata.
• A Python loader, validation tools, and PE/NVS task converters.

Each MP4 stacks the left eye above the right eye. Each eye is 1280×1024 pixels; sessions 001–002 use 30 FPS and sessions 003–007 use 60 FPS. V2 pose tables store row-major 4×4 transforms from endoscope1 left to endoscope2 left, with translation in metres. Per-sequence metadata records pose availability and valid frame ranges. All files retain their released contents, including video resolution and encoding.

The v1.0 sample release contains one scene per session (~6 GB). The loader supports both versions.

Getting started

The viewer shows one row per recording with both endoscopes and sequence metadata. It displays processed videos where available and raw videos for calibration recordings. The recordings split contains the complete release. The first 100 rows mix camera movements, stationary views, tool interactions, and calibration recordings. H.264 browser copies in previews/ keep the original resolution and frame rate. Dataset downloads use the original files in imed_v2_0/.

Download the full dataset (~48 GB) with the Hugging Face CLI:

python -m pip install -U huggingface_hub
hf download sbonilla/imed --repo-type dataset --include "imed_v2_0/**" --local-dir .

Select a session, scene, recording, or file type with --include:

# Ex-vivo sessions, including camera calibration
hf download sbonilla/imed --repo-type dataset --local-dir . \
  --include "imed_v2_0/session_00[123]/**" \
  --include "imed_v2_0/calibrations/session_00[123]/**" \
  --include "imed_v2_0/dataset_metadata.json"

# One scene, including metadata and calibration for the loader
hf download sbonilla/imed --repo-type dataset --local-dir . \
  --include "imed_v2_0/session_007/scene_9/**" \
  --include "imed_v2_0/session_007/session_metadata.json" \
  --include "imed_v2_0/calibrations/session_007/**" \
  --include "imed_v2_0/dataset_metadata.json"

For post-mortem sessions, use session_00[456] in both patterns; for live porcine, use session_007.

For one recording, narrow the scene pattern to imed_v2_0/session_007/scene_9/circular/**. For raw videos only, use --include "imed_v2_0/**/*_raw.mp4". The UCL dataset record also provides the ZIP.

Filter by metadata

Each session_metadata.json lists its tissue type and recording labels, including movement_type, tools_present, calibration_sequence, and pose_type. Read these small files first, then download the matching recording folders.

Python example: select tissue type and camera motion

Set tissue to ex_vivo, in_vivo_postmortem, or in_vivo_live (or None for all). Set selection to moving, stationary, or calibration. Here, stationary selects recordings with static camera poses outside the calibration sequences.

import json
from pathlib import Path
from huggingface_hub import HfApi, hf_hub_download, snapshot_download

tissue = "ex_vivo"
selection = "moving"
repo = "sbonilla/imed"
options = dict(repo_id=repo, repo_type="dataset", local_dir=".",
               revision=HfApi().dataset_info(repo).sha)
moving = {"circular_camera_motion", "lateral_camera_motion", "zoom_camera_motion"}
filters = {
    "moving": lambda seq: seq["movement_type"] in moving,
    "stationary": lambda seq: seq.get("pose_type") == "static"
                              and not seq["calibration_sequence"],
    "calibration": lambda seq: seq["calibration_sequence"],
}
matches = []
patterns = {"imed_v2_0/dataset_metadata.json"}
for number in range(1, 8):
    session = f"session_{number:03d}"
    meta = f"imed_v2_0/{session}/session_metadata.json"
    path = hf_hub_download(filename=meta, **options)
    data = json.loads(Path(path).read_text())
    if tissue is not None and data["tissue_type"] != tissue:
        continue
    for region in data["anatomical_regions"]:
        for seq in region["sequences"]:
            if filters[selection](seq):
                matches.append(seq["sequence_id"])
                patterns.update({f"imed_v2_0/{seq['sequence_id']}/**", meta,
                                 f"imed_v2_0/calibrations/{session}/**"})

if not matches:
    raise ValueError("No recordings match these filters")
print(f"Downloading {len(matches)} recordings")
snapshot_download(allow_patterns=sorted(patterns), **options)

Each selected folder includes its videos, timestamps, poses, and masks where present. Add conditions such as and seq["tools_present"] to select tool recordings.

Load recordings

Place the scripts repository alongside imed_v2_0/:

git clone https://github.com/surgical-vision/imed.git imed_scripts
python -m pip install -r imed_scripts/requirements.txt
python imed_scripts/examples/list_dataset.py imed_v2_0

Requires Python 3.11 or newer. See the scripts README for frame loading, pose queries, task conversion, and stereo depth generation.

Challenge

The iMED challenge includes pose estimation (iMED-PE) and deformable novel view synthesis (iMED-NVS):

• Challenge website
• Synapse challenge page

Citation and licence

Please cite the paper and the dataset record when using iMED. The dataset is released under CC BY 4.0.

@inproceedings{bonilla2026imed,
  title     = {{iMED}: A Multi-endoscope Dataset for Surgical 3D Perception},
  author    = {Bonilla, Sierra and Jiang, Fengyi and Nwoye, Chinedu and Lu, Jingpei and Reardon, Kailey and Chow, Humphrey W. and Vasconcelos, Francisco and Bano, Sophia and Schmidt, Adam and Mohareri, Omid},
  booktitle = {Computer Vision -- ECCV 2026},
  pages     = {313--332},
  year      = {2026},
  doi       = {10.1007/978-3-032-37393-9_19}
}