SurgBox contains annotations only. The Cholec80 videos are distributed by CAMMA under their own terms, and SurgBox may be used for non-commercial purposes only (CC BY-NC-SA 4.0).

SurgBox

Instance-level bounding boxes for surgical instruments in all 80 videos of Cholec80.

Paper: From Presence Labels to Bounding Boxes: Can Multimodal Large Language Models Scale Surgical Instrument Localization? (MICAD 2026)
Code: github.com/M-Hamdy-M/SurgBox
Models: SurgBox-YOLOv8-x, SurgBox-RT-DETR-L, SurgBox-YOLO11-x (paper, Table 2) and SurgBox-Cross-Dataset (Table 3)

Cholec80 annotates, once per second, which of seven instruments are present, but not where they are. SurgBox converts these presence labels into class-specific bounding boxes using a multi-stage multimodal LLM pipeline, and provides a 1,000-frame human reference, annotated twice and reviewed by a surgeon, for measuring localization quality.

This repository contains annotations only. The videos must be requested from CAMMA.

Configurations

Config	Split	Videos	Frames	Boxes
`surgbox`	train	56	128,172	201,596
	validation	8	16,971	25,428
	test	16	39,355	58,953
`human_reference`	test	80	1,000	1,313

Boxes per class in surgbox:

Grasper	Bipolar	Hook	Scissors	Clipper	Irrigator	SpecimenBag
139,464	9,010	103,495	4,164	6,708	11,295	11,841

from datasets import load_dataset

surgbox = load_dataset("M-Hamdy/SurgBox")                                 # train, validation, test
human = load_dataset("M-Hamdy/SurgBox", "human_reference", split="test")

Loading requires datasets 4.0 or later. Accept the conditions on this page and log in with hf auth login first.

Fields

Each row is one frame.

Field	Description
`image_id`	image id in `coco/surgbox.json.gz` (`surgbox`) or `coco/human/*.json` (`human_reference`); the two differ, so join the configs on `video` and `frame`
`file_name`	`video01/000025.png`, the frame's path after extraction (below)
`video`	Cholec80 video number, 1 to 80
`frame`	frame index in the original 25 fps video, matching the `Frame` column of the Cholec80 tool annotations
`width`, `height`	854×480, except videos 78, 79 and 80 (1920×1080)
`objects`	`category`, `bbox`, `area` and `confidence` of each box (`surgbox` only)
`reference`, `annotator_1`, `annotator_2`	`category`, `bbox` and `area` of each box (`human_reference` only)
`review`	surgeon review of the reference frame (`human_reference` only)

Frames are sampled at 1 fps. Boxes are [x, y, width, height] in pixels of the source video. A box covers the visible extent of the instrument, shaft and tip, without completing occluded or out-of-frame parts. Some boxes (423 of 285,977) extend past the image border; before training, clip boxes to the image and drop any side shorter than 2 pixels, as in the detection benchmark of the paper. confidence (1–5) is the MLLM's own confidence in each box. category is one of Grasper, Bipolar, Hook, Scissors, Clipper, Irrigator and SpecimenBag, stored as a class index (0–6) in the Parquet files and as category_id (1–7) in the COCO files.

The same annotations are also provided in COCO format:

coco/
  surgbox.json.gz         all 184,498 frames
  human/
    reference.json        annotator 1 after surgeon review (the reference)
    annotator_1.json      annotator 1 before review
    annotator_2.json      annotator 2, independent pass
  splits.json             video-level splits used in the paper

To extract the annotated frames from a Cholec80 video:

import cv2, os

def extract(video, out_dir):
    os.makedirs(out_dir, exist_ok=True)
    cap, i = cv2.VideoCapture(video), 0
    while (ok := cap.read())[0]:
        if i % 25 == 0:
            cv2.imwrite(f"{out_dir}/{i:06d}.png", ok[1])
        i += 1

extract("video01.mp4", "video01")

Human reference

The 1,000 frames span all 80 videos and are sampled from the Cholec80 labels to balance the seven classes, with 150 frames that have no instrument in the labels. Each frame was annotated by two annotators; the first annotation set was reviewed by a surgeon. review is finalized for the 984 frames the surgeon approved and changes_requested for 16 frames outside the laparoscopic view (e.g. out-of-body views), which have no boxes.

	Set F1 (%)	mIoU	Matched mIoU
Annotator 2 vs. reference (human agreement)	94.18	0.814	0.912
Pipeline, evaluation run (paper, Table 1)	94.66	0.784	0.857
`surgbox` (this release)	94.42	0.774	0.852

The released annotations are a separate run of the same pipeline over all frames; the pipeline is stochastic, so they score slightly differently on the reference than the evaluation run reported in the paper. Boxes are matched greedily within each class by IoU, and an unmatched reference box counts as zero. The evaluation script is in the GitHub repository.

The reference frames come from all 80 videos, 702 of them from training videos. To evaluate a detector trained on the train split, use the 197 frames from test videos with review = finalized (238 boxes), the expert-verified subset in the paper (Table 2).

Splits

The surgbox splits are video-disjoint (56 / 8 / 16 videos). coco/splits.json also lists the five CholecSeg8k test videos (28, 43, 48, 52, 55) used in the cross-dataset experiment; four of them are in train and were excluded from training in that experiment. The detection benchmark in the paper omits one test frame (video 44, frame 69,525), so its test split has 39,354 frames instead of 39,355.

Citation

If you use SurgBox, please cite our paper:

@inproceedings{hamdy2026surgbox,
  title     = {From Presence Labels to Bounding Boxes: Can Multimodal Large Language Models Scale Surgical Instrument Localization?},
  author    = {Hamdy, Mohamed and Abdel-Ghani, Muraam and Ahmed, Fatmaelzahraa and Nasar, Sifna and Ahmed, Mariam and Al-Jalham, Khalid and Al-Ali, Abdulaziz and Balakrishnan, Shidin},
  booktitle = {Medical Imaging and Computer-Aided Diagnosis (MICAD)},
  year      = {2026}
}

Source datasets (please cite these in addition to SurgBox)

SurgBox is built on the Cholec80 videos and instrument-presence labels; the cross-dataset experiment uses CholecSeg8k.

@article{twinanda2017endonet,
  title   = {EndoNet: A Deep Architecture for Recognition Tasks on Laparoscopic Videos},
  author  = {Twinanda, Andru P. and Shehata, Sherif and Mutter, Didier and Marescaux, Jacques and de Mathelin, Michel and Padoy, Nicolas},
  journal = {IEEE Transactions on Medical Imaging},
  volume  = {36},
  number  = {1},
  pages   = {86--97},
  year    = {2017}
}

@article{hong2020cholecseg8k,
  title   = {CholecSeg8k: A Semantic Segmentation Dataset for Laparoscopic Cholecystectomy Based on Cholec80},
  author  = {Hong, W.-Y. and Kao, C.-L. and Kuo, Y.-H. and Wang, J.-R. and Chang, W.-L. and Shih, C.-S.},
  journal = {arXiv preprint arXiv:2012.12453},
  year    = {2020}
}

License

The annotations are released for non-commercial use only, under CC BY-NC-SA 4.0, consistent with Cholec80 and CholecSeg8k. Cholec80 is not redistributed and remains subject to its own terms.