oluwabukolaadegboro/instance_segmentation_with_mask_rcnn

Mask_RCNN

This also consitutes a part of my computer vision coursework taught by Georgia Gkioxari (a research scientist at Facebook) during my African Masters in Machine Intelligence (AMMI) program at AIMS Rwanda.

In this project, a nuts custom dataset was prepared and trained on an object detection library (Detectron2). The nuts dataset had 3 classes: date, fig and hazelnut, all annotated with instance masks. The dataset can be gotten here. The data contained an image directory which has the images, a train.json file that contains the train annotations in COCO format, and a val.json file that contains the val annotations in COCO format.

After preparing and registering the data, it was with trained a Mask R-CNN model with a ResNet50 feature Pyramid Network (FPN) backbone. I initialized the model with two initialization schemes: a scheme with an existing model pre-trained on the COCO dataset and a scheme with ImageNet weight. Finally, I trained for 300 iterations using a start learning rate of 0.02, 2 images per batch, and 128 regions per batch. Below are some of my visualizations I came across during the implementation of this work.

Visualization of the both predictions for CocoNet initialization (right) and ImageNet initialization (left) on the validation set

CocoNet Initialization AP Evaluation

ImageNet Initialization AP Evaluation