Instance-segmentation-with-Mask-RCNN

This is the unofficial implementation of the paper Mask-RCNN using pytorch-lightning. We demonstrate Mask RCNN for a subset of COCO dataset - which segments three categories namely people, animals and vehicles.

Downoad dataset and required modules by running

!gdown 101kiCyfTeGvs1p5nW3UxH9oZJ8ucJ3hx
!gdown 1fnbbHeZ3DZqsXwPDrq9YF3p8vI3PCK1K
!gdown 1YCK8Sfbj_UDp1mnSA6g55hlfcpcSklbv
!gdown 1KIf6jeMPpfvWqqiGqjc6Y4JmGHkbL9_C
!gdown 1h6VQWmbq41cJ9O1WdRalc8iOsox2HCeO 

Setup instructions

It is recommended to create a virtual environment and run

pip install -r requirements.txt

Running the code

python main.py

It is important to note that the main function in main.py, trains three modules sequentially (RPNHead ,BoxHead, and MaskHead) -- the different training instances are specified with appropriate comments. For better results it is recommended to train all three modules together after the provided sequential training regime.