Mask RCNN in TensorFlow
This repo attempts to reporduce this amazing work by Kaiming He.
Mask RCNN.
- Allocate-Zero-Memory Error, solved by @santisy
- Allowing larger initial learning rate.
- Chooseing proper initial weights.
- Download coco dataset, place it into
./data
, then runpython download_and_convert_data.py
to build tf-record. It takes a while. - Download pretrained resnet50 model,
wget http://download.tensorflow.org/models/resnet_v1_50_2016_08_28.tar.gz
, unzip it, place it into./data/pretrained_models/
- run
python test/resnet50_test.py
for training - There are certainly some bugs, please report them back, and let's solve them togather.
- ROIAlign
- COCO Data Provider
- Resnet50
- Feature Pyramid Network
- Anchor and ROI layer
- Mask layer
- Speedup anchor layer with cython
- Combining all modules together.
- Testing and debugging (in progress)
- Training / evaluation on COCO
- Data queue
- Data agument
- Other backbone networks
- Training >2 images
- Anything helps this repo, including discussion, testing, promotion and of course your awesome code.