Keras Mask R-CNN for Open Images Challenge 2019: Instance Segmentation

Repository contains Mask R-CNN models which were trained on Open Images Dataset during Kaggle competition: https://www.kaggle.com/c/open-images-2019-instance-segmentation/leaderboard

Repository contains the following:

• Pre-trained Mask R-CNN models (ResNet50, ResNet101 and ResNet152 backbones)
• Example code to get predictions with these models for any set of images
• Code to train model based on Keras Mask R-CNN and OID dataset

Requirements

Python 3.*, Keras 2.*, keras-maskrcnn 0.2.2, cv2, numpy, pandas

Pretrained models

There are 3 Mask R-CNN models based on ResNet50, ResNet101 and ResNet152 for 300 classes.

Backbone	Image Size (px)	Model	Small validation mAP	LB (Public)
ResNet50	800 - 1024	521 MB	0.5745	0.4259
ResNet101	800 - 1024	739 MB	0.5917	0.4345
ResNet152	800 - 1024	918 MB	0.5899	0.4404

• Model - can be used to resume training or can be used as pretrain for your own instance segmentation model

Inference

Simple example can be found here: inference_example.py

Training

For training you need to download OID dataset (~500 GB images): https://storage.googleapis.com/openimages/web/download.html You need all images, all masks and all CSV-files related to Instance Segmentation track.

Then run script (change parameters and file locations at the bottom of script):

• training/train_maskrcnn.py