Step2_DetectCrop/README.md

Step2 OD/Fovea Detection

This directory contains the coordinate-detection stage. It outputs OD/Fovea coordinates and does not generate crop boxes directly.

Model weights:

• weights/RetinaNet_cfp_v.pth (wfCFP + vessel-enhanced fusion input)
• weights/RetinaNet_octa_v.pth (OCTA + vessel-enhanced fusion input)

Single-Pair Run

python run_detect_crop.py \
  --wfcfp_image path/to/wfcfp.png \
  --wfcfp_vessel path/to/wfcfp_vessel.png \
  --octa_image path/to/octa.png \
  --octa_vessel path/to/octa_vessel.png \
  --output_dir Output/demo

Main outputs:

• wfcfp_od_fovea.txt
• octa_od_fovea.txt (when OCTA input is provided)
• summary.json

OCTA-60 Batch Run

bash scripts/run.sh

Notes:

• run_detect_crop.py keeps --crop_output_size only for backward compatibility.
• Final coordinate correction follows the logic in detect_server.py.