2018 Data Science Bowl 2nd Place Solution
My solution is a modification of Unet. To make Unet instance‐aware, I add eight more outputs describing the relative positions of each pixel within every instance as shown in the images below. In my final model version, the entire network structure of Unet before the output layers was replaced by the pre‐trained Mask‐RCNN feature extractor (P2 as in Matterport version of MaskRCNN) for better performance.
https://github.com/jacobkie/2018DSB/blob/master/imgs/0.png
The relative position masks are shown as below: https://github.com/jacobkie/2018DSB/blob/master/imgs/1.png https://github.com/jacobkie/2018DSB/blob/master/imgs/1.png
Codes in utils.py, parallel_model.py, params.py, visualize.py, model_rcnn_weight.py are partly adapted from Matterport Mask_RCNN (https://github.com/matterport/Mask_RCNN) which is under MIT license. I also used its pre‐ trained weight on MS COCO (https://github.com/matterport/Mask_RCNN/releases).
Four sources of data were used:
Some masks of the 2009ISBI data set are manually modified.
To train from scratch
Or you can use my pretrained weight in the cache folder to make predictions directly.