Counting 3,742 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1


Faster Training of Mask R-CNN by Focusing on Instance Boundaries

Instance Mask Visualizations

This is an implementation of the improved training scheme Faster Training of Mask R-CNN by Focusing on Instance Boundaries on Python 3, Keras, and TensorFlow. The code is an extension of the existing implementation of Mask R-CNN by Matterport. It can be seen as a fork of the original repository based on commit cbff80f. The model generates bounding boxes and segmentation masks for each instance of an object in the image. It's based on Feature Pyramid Network (FPN) and a ResNet101 backbone. The training speed has been increased by introducing an auxiliary objective.

Adding the new auxillary task can be done using this simple network head:


The choice of the edge detection filter influences the convergence speed up; as written in the paper, the best results were obtained using the Sobel filter.

Loss curves

If you like this work and want to use this in your work or research, please cite:
author    = {Roland S. Zimmermann and
            Julien N. Siems},
title     = {Faster Training of Mask {R-CNN} by Focusing on Instance Boundaries},
journal   = {CoRR},
volume    = {abs/1809.07069},
year      = {2018},
url       = {},
archivePrefix = {arXiv},
eprint    = {1809.07069},
timestamp = {Fri, 05 Oct 2018 11:34:52 +0200},
biburl    = {},
bibsource = {dblp computer science bibliography,}