This project is a little experimental implementation of the Mask R-CNN algorithm for image segmentation. As for the Experiment, I wanted to find out whether the following priciple actually holds:
Given a backbone that performs well on classifying my objects
(on single-object images), that backbone is also suitable for
detecting and masking my objects (on multi-object images).
My limited resources inspired me to try this principle with a "minimal" example, i.e. the MNIST dataset. Therefore, the goal of this project is detection and masking of scattered MNIST-letters.
- Most of the model implementation is inspired by the keras implementation by Matterport.
- The MNIST backbone was developed based on the keras MNIST example.
- Crop and mask images:
crop_and_mask_image_files()
- Load data like the mnist-module:
load_data()
- Generate random images with labels:
generate_labeled_data_files()
- Load generated data:
load_labeled_data()
- Parse data:
load_backbone_pretraining_data()
,load_maskrcnn_data()
- Inspect data:
write_solutions()