FastMaskRCNN_ForwardTest

This is a forward test script of a single input image for the FastMaskRCNN: https://github.com/CharlesShang/FastMaskRCNN

Requirements

• Tensorflow (>= 1.0.0)

• Numpy

• PIL (Pillow version = 2.3.0)

• Pre-trained Resnet50

• Pre-trained FastMaskRCNN model (From: FastMaskRCNN #107)

Functionalities

• Draw bounding box of the predicted RoI

• Draw mask of the predicted RoI (without bbox)

• Draw mask of the predicted RoI with its bbox and predicted categroy

Getting Start

1. It requires you to download the whole repo firstly from https://github.com/CharlesShang/FastMaskRCNN

2. Add the whole dir ./forward_test under the root of the repo; then replace the original ./libs/visualization/pil_utils.py with the new one in my repo. Downloading the 2 pre-trained models above and place them(with the ./output/mask_rcnn/checkpoint) as shown below. Finally the folder structure would be:

root/
├── data/pretrained_models/resnet_v1_50.ckpt
│   └── ......
│
├── forward_test/
│   ├── testdata/
│   ├── output/
│   └── forward_test_single_image.py
│
├── libs/
│   ├── visualization/
│   │   ├── pil_utils.py
│   │   └── ......
│   ├── nets/
│   │   ├── pyramid_network.py
│   │   └── ......
│   └── ......
│
├── output/mask_rcnn/
│   ├── checkpoint
│   ├── coco_resnet50_model.ckpt-2499999.data-00000-of-00001
│   ├── coco_resnet50_model.ckpt-2499999.index
│   └── coco_resnet50_model.ckpt-2499999.meta
└── ......

3. Modify original ./libs/nets/pyramid_network.py according to Issues#1-F3 and Issues#1-F4. Or you can just replace the original one with mine.

4. Put your test image under ./forward_test/testdata/

5. (Optional) If you want to change the output image dir, modify the code ./forward_test/forward_test_single_image.py at Line30-31

save_dir_bbox = 'output/bbox/'
save_dir_mask = 'output/mask/'

6. (Optional) If your test image is in PNG format, modify the code ./forward_test/forward_test_single_image.py at Line32

file_pattern = 'jpg'    # or 'png'

7. run ./forward_test/forward_test_single_image.py and wait for the result

Acknowledgment

• The ./forward_test/forward_test_single_image.py is modified from the original ./train/train.py from FastMaskRCNN

• The ./libs/visualization/pil_utils.py is modified from @chen1005's suggestion

• The pre-trained Mask-RCNN model from @QtSignalProcessing

To-Do

• Fix the bug of Issues#2

• Add DenseCRF