• train keras-retinanet with modification to allow configuration of NMS from cli
  • that change is here https://github.com/sshleifer/keras-retinanet/commit/2f9bd031edb76348a00d97d938811d188833da58
• experiment with params: 256x256 with batch_size=12, coco starting weights, resnet50
• run inference with nms_threshold=.01
• calc each images probability as the highest bbox probability
• only submit detections for the 360 test images with the highest image proba
• see retinanet_inference.py

train mask-rcnn train keras-retinanet on 256x256 train keras-retinanet on 300x300 no NIH data inference: retinanet_inference.py use nms_threshold=0.01 (combine boxes if there is any overlap) use score_threshold=.15 for retinanet, .95 for mask-rcnn

Optionally: can also ensemble resnet101 backed mask-rcnn

This repo is mostly modifications to the mask-rcnn code and the utils for ensembling. The ensembling didn't work.

see thrensemble(): never combines detections across models, just chooses which models to use