runist / fcn-keras

Using keras and tf build FCN

Keras - FCN

Part 1. Introduction

Fully Convolutional Networks is the first model to apply Convolutional Neural Network to semantic segmentation. It used common backbone like VGG, ResNet as encoder, and the decoders are upsampled layer by layer to original image size.

Models for this repository

Model name	Dataset	MIoU	Pixel accuracy
FCN_8_Resnet50	VOC train dataset	0.4578	0.8993
	VOC validation dataset	0.751	0.9528

Part 2. Quick Start

1. Pull this repository.

git clone https://github.com/Runist/FCN-keras.git

2. You need to install some dependency package.

cd FCN-keras
pip install -r requirements.txt

3. Download the VOC dataset(VOC SegmetationClassAug if you need) .
4. Getting FCN weights.

wget https://github.com/Runist/FCN-keras/releases/download/v0.2/fcn_weights.h5

4. Run predict.py, you'll see the result of Fully Convolutional Networks.

python predict.py

Input image:

Output image（resize to 320 x 320）:

Part 3. Train your own dataset

1. You should rewrite your data pipeline, Dateset where in dataset.py is the base class, such as VOCdataset.py.

class VOCDataset(Dataset):
    def __init__(self, annotation_path, batch_size=4, target_size=(320, 320), num_classes=21, aug=False):
        super().__init__(target_size, num_classes)
        self.batch_size = batch_size
        self.target_size = target_size
        self.num_classes = num_classes
        self.annotation_path = annotation_path
        self.aug = aug
        self.read_voc_annotation()
        self.set_image_info()

2. Start training.

python train.py

3. Running evaluate.py to get mean iou and pixel accuracy.

python evaluate.py
--------------------------------------------------------------------------------
Total MIOU: 0.4578
Object MIOU: 0.4357
pixel acc: 0.8993
IOU:  [0.89910368 0.57704284 0.42540458 0.37512895 0.20381801 0.35467035
 0.77999658 0.63954053 0.6688376  0.22855226 0.36980252 0.25836402
 0.59642955 0.48159056 0.42752974 0.67481864 0.13373938 0.33782181
 0.30385944 0.47929764 0.39788109]

Part 4. Paper and other implement

• Fully Convolutional Networks for Semantic Segmentation
• paper with code: shelhamer/fcn.berkeleyvision.org
• aurora95/Keras-FCN
• divamgupta/image-segmentation-keras