GDXray-retinanet

This project applies the RetinaNet object detector on the GDXray dataset, Castings group.

Reuse

We entirely reuse the Keras Retinanet Object Detection Framework and just apply the dataset as instructed by the framework.

Configuration

We use the frameworks's default Focal Loss hyper-paramters of alpha=0.25, and gamma=2.0

Training

./train.py --multi-gpu=3 --batch-size=3 --freeze-backbone \
--no-evaluation --steps=10000 --epochs=20 --snapshot-path xray3-snapshots csv ../\
utils/annotations-with-negatives/train_annotations.csv ../utils/annotations/classes.csv

Resume training from a snapshot (--snapshot)

./train.py --gpu=1 --freeze-backbone --no-evaluation --steps=10000 --epochs=5 \
--snapshot  xray3-snapshots/resnet50_csv_20.h5 \
--snapshot-path xray4-snapshots \
csv ../utils/annotations-with-negatives/train_annotations.csv ../utils/annotations/classes.csv

Evaluate

./evaluate.py --save-path results/ --max-detections=7 \
csv ../utils/annotations-with-negatives/test_annotations.csv \
../utils/annotations-with-negatives/classes.csv xray-snapshots/resnet50_csv_11.h5

Initial Results

With the default Focal Loss hyperparamter settings, the mAP is 0.76

A sampling of inital results show below.

Ground-truth bounding-boxes are in green, detections are in blue.