README.md

Mask-RCNN-pytorch

Pytorch implementation of Mask-RCNN based on torchvision model with VOC dataset format. The model generates segmentation masks and their scores for each instance of an object in the image. This repository is based on TorchVision Object Detection Finetuning Tutorial.

Training

label your data with labelme and Export VOC-format dataset from json files with labelme2voc.

Prepare your dataset in this format:

my_dataset
      ├── labels.txt
      │
      ├── JPEGImages
      │       ├── image1.jpg
      │       └── image2.jpg
      │
      ├── SegmentationObject
      │       ├── image1.png
      │       └── image2.png
      │
      └── SegmentationClass
              ├── image1.png
              └── image2.png

Clone the repository and put my_dataset folder in Mask-RCNN-pytorch folder then use this line of code to train:

$ python3 train.py --data my_dataset --num_classes 11 --num_epochs 150

Enter num_classes including background.

Testing

Enter your class names using classes variable in mask_rcnn.py then use this line of code to test on your image:

$ python3 test.py --img test_img.jpg --model ./maskrcnn_saved_models/mask_rcnn_model.pt

Here are some output results: