tkwoo / visualization-segmentation-process

Segmentation visualization, keras, augmentation, fine tuning

Visualization segmentation training process

Semenatic segmentation using Unet, fcn, pspnet

Result

click image to watch video

Requirements

• Python 2.7 or 3.6
• OpenCV 3.4.0
• Keras 2.1.4
• TensorFlow 1.5.0

Usage

To train a model (visualization)

$ python main.py

Then, the training steps(image) will be saved 'result' directory

---

usage: main.py [-h] [--data_path DATA_PATH] 
                    [--output_dir OUTPUT_DIR]
                    [--image_height IMAGE_HEIGHT] 
                    [--image_width IMAGE_WIDTH]
                    [--batch_size BATCH_SIZE]
                    [--total_epoch TOTAL_EPOCH]
                    [--initial_learning_rate INITIAL_LEARNING_RATE]
                    [--learning_rate_decay_factor LEARNING_RATE_DECAY_FACTOR]
                    [--epoch_per_decay EPOCH_PER_DECAY] 
                    [--ckpt_dir CKPT_DIR]
                    [--ckpt_name CKPT_NAME]
                    [--pretrained_weight_path PRETRAINED_WEIGHT_PATH]
                    [--confidence_value CONFIDENCE_VALUE] 
                    [--debug DEBUG]
                    [--mode MODE] 
                    [--test_image_path TEST_IMAGE_PATH]
                    [--tf_log_level TF_LOG_LEVEL]

Input data(only for training)

└── dataset
    └── xxx
        └── train
            └── IMAGE
                └── ori
                    └── xxx.png (name doesn't matter)
            └── GT
                └── mask
                    └── xxx.png (It must have same name as original image)

The dataset directory structure is quite complex to use the Keras ImageDataGenerator Framework.

Input data for testing

└── test_data
    └── image.png

First, create checkpoint dir and download trained parameter files

└── checkpoint
    └── (ckpt_name)
        ├── model.json 
        ├── weight.xx.h5
        └── ...

You can download CHECKPOINT --> not supported

To test a model

$ python main.py --mode predict_img --ckpt_name <NAME> --test_image_path <.../image.png>

Reference

paper : https://lmb.informatik.uni-freiburg.de/people/ronneber/u-net/

visualization-segmentation-process's Issues

Pretrained model

Can you please share the pretrained model?

About predict.py

This visualization tool is interesting.
I would like to ask you
What is specified for the test image format etc?
ｐredict.py can not be executed due to an error.

input_data = imgInput.reshape((1,256,256,1))
AttributeError: 'NoneType' object has no attribute 'reshape'

I am adjusting the image conditions, but what do you suppose to do with formats and image matrices?

Wrong regressor values

I guess the color values are wrong. Correct me if Im mistaken

    def make_regressor_label(self, gt):
        human = np.where(gt==24,255,0) + np.where(gt==25,255,0)
        car = np.where(gt==26,255,0) + np.where(gt==27,255,0) + np.where(gt==28,20,0)
        road = np.where(gt==7,255,0) #+ np.where(gt==8,1,0)
        label = np.concatenate((human, car, road), axis=-1)

        return label

The color code for the human in the segmentation mask is #181818 which in rgb is (24,24,24) but we are searching for wrong values between (24,255,0) and (24,255,0). why so?