MEV Annotation Toolkit

This directory contains the code used to generate multi-event video annotations.

Layout

qwen/: local Qwen3-VL annotation, event segmentation, and camera-motion classification scripts.
gemini/: Gemini API wrappers that reuse the Qwen annotation pipeline schema and post-processing.
scripts/multi-event/: shell launchers for batched multi-GPU annotation runs.

Generated annotations, logs, checkpoints, and datasets are intentionally not included. Keep them under local runtime directories such as data/, annotations/, logs/, and checkpoints/.

Basic Usage

Run Qwen annotation on segmented videos:

bash scripts/multi-event/run_multi_event_annotations.sh data/example_segments annotations/qwen/multi_event/output.json --resume

Run Gemini annotation:

export GEMINI_API_KEY=your_api_key
python gemini/generate_multi_event_annotations.py --input data/example_segments --output annotations/gemini/multi_event/output.json --recursive --resume

Segment a single video into event intervals:

python qwen/segment_events_qwen3vl.py data/video.mp4 --output annotations/segments/video_events.json

Classify camera motion for a single video:

python qwen/classify_camera_motion_qwen3vl.py data/video.mp4 --output annotations/camera/video_camera.json

Notes

The Qwen scripts expect a local model checkpoint via --model-path or MODEL_PATH.
The shell launchers derive all paths from this directory and environment variables.
ffmpeg and ffprobe must be available on PATH.
Gemini scripts use GEMINI_API_KEY or GOOGLE_API_KEY.