This is a Pytorch 1.0 implementation of Mask R-CNN that is based on Matterport's Mask_RCNN[1] and this[2]. Matterport's repository is an implementation on Keras and TensorFlow.
The main improvements from [2] are:
- Pytorch 1.0
- most numpy computations were ported to pytorch (for GPU speed)
- supports batchsize > 1
- some bugs were fixed in the translation process
- code refactor
- NMS speed-up
Currently, it works with Kaggle's 2018 Data Science Bowl dataset (the result on 1st phase testset is 0.27).
to train the network use: python samples/nuclei.py train --dataset=path_to_dataset --model=coco
to detect use: python samples/nuclei.py submit --dataset=path_to_dataset --model=path_to_trained_model
to check Kaggle's 2018 Databowl metric on a dataset use: python samples/nuclei.py metric --dataset=path_to_dataset --model=path_to_trained_model
for installation instructions, just export mrcnn directory to PYTHONPATH and run: python setup.py install