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
Python 3.*, Keras 2.*, keras-maskrcnn 0.2.2, cv2, numpy, pandas
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
Simple example can be found here: inference_example.py
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):