EndoComparer / README.md
ethanshili's picture
EndoComparer release
635800e
|
Raw History Blame Contribute Delete
2.14 kB
metadata
license: cc-by-nc-sa-4.0
base_model: Qwen/Qwen3.5-9B
library_name: peft
pipeline_tag: video-text-to-text

EndoComparer

EndoComparer is a vision-language model focused on cross-case surgical video comparison: given two surgical videos from different cases, it answers questions about whether and how they differ in anatomy, tools, actions and skill. It gives each video its own visual encoding and reasoning path, which makes the model more robust and accurate against video-binding confusion, where the content of one video is attributed to the other.

πŸš€ Quick Start

1. Install

git clone https://github.com/CAMMA-public/endocomparer.git
cd endocomparer
pip install -r requirements.txt

2. Get the base model and the EndoComparer checkpoint

hf download Qwen/Qwen3.5-9B --local-dir models/Qwen3.5-9B
hf download ethanshili/EndoComparer --local-dir models/EndoComparer

3. Run inference on the demo item

demo/ ships one EndoDiffVQA test item: two cholecystectomy clips and the question "Is the gallbladder coagulated in both videos?" with its four options.

bash scripts/evaluate.sh --jsonl demo/items.jsonl \
    --model models/Qwen3.5-9B --weights models/EndoComparer \
    --out results/demo/results.json

The model first states what each video shows, then gives the letter:

Video 1: coagulated. Video 2: not coagulated. (b)

Please refer to our GitHub repository for training and evaluation.

πŸ“– Citation

To be added.