- train keras-retinanet with modification to allow configuration of NMS from cli
- 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