Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
These two clips are taken from HeiCo-FOCUS-VQA, part of the ORena FOCUS Challenge (https://orena-focus-challenge.org), and are shared with the organisers' agreement. By downloading them you agree to the HeiCo-FOCUS data usage terms: (1) You may not use the data, any algorithms trained on it, or any results of such algorithms for publication purposes before the first version of the Challenge publication has been submitted. (2) Passing on data, parts of it, or any derivatives can only be done, once the recipient also accepts clause (1).
Log in or Sign Up to review the conditions and access this dataset content.
SurgZoom examples (HeiCo-FOCUS-VQA)
Two question–answer cases from the test split of HeiCo-FOCUS-VQA, prepared for the quick start of SurgZoom (model: wxyi088/orena-SurgZoom). The folder uses the input layout of the ORena SAVE FOCUS challenge container, so the same command runs locally and on the platform.
| qID | Question | Clip (source timeline) | Answer | What it shows |
|---|---|---|---|---|
| 2616489 | There is one Sponge in the frame at 00:51:42. When is it retrieved from the surgical site? Please provide the answer in hh:mm:ss. | 00:50:31 – 00:52:30 | 00:52:05 | anchored sub-window + temporal zoom |
| 2143804 | What is the maximum number of Clips appearing at once in a single frame? Please provide a number. | 00:15:48 – 00:17:47 | 3 | dense sampling for counting |
Both clips come from sigmoid resections.
Files
request.json questions in the challenge request format
batch.json batch description (challenge format)
FO_definitions.json foreign-object class definitions (a JSON string), used in the system prompt
overlayed/<qID>.mp4 question clips: 5 fps, 960x540, H.264, absolute HH:MM:SS burned in top-left
anchored/2616489.mp4 the anchored sub-window of case 1 (00:51:32 - 00:52:30), used for training
reference.json ground truth rows with the original dataset columns
train_smoke.jsonl the two cases as ms-swift training rows (regime III), for a smoke test
Clips were cut from the original HeiCo videos with tools/prepare_data.py of the SurgZoom
repository (the recipe used to build its training data). Clip file names are the dataset
question ids.
Use
git clone https://github.com/wxyi057/orena-SurgZoom && cd orena-SurgZoom && pip install -e ".[eval]"
bash scripts/download_examples.sh # this dataset -> examples/heico
python -m surgzoom.infer --input examples/heico --output answers.json
python tools/evaluate.py --answers answers.json --reference examples/heico/reference.json --judge none
Licence and terms
CC BY-NC-SA 4.0, following HeiCo-FOCUS-VQA, plus the data usage terms shown above (accepted when requesting access). The full dataset is available from orena-dkfz/heico-focus-vqa.
- Downloads last month
- 14