Construct a automatically training and testing framework to evaluate the performance of the model. This framework includes the following modules.

1. preprocessing
   1. random cropping
   2. image augmentation using imgaug
2. model construction
   1. unet
   2. deeplab with weights pretrained on pascal VOC
   3. mrcnn
3. evaluation
   1. official pycocotools

1. opencv 4.0.0
2. imgaug 0.2.6
3. keras 2.2.4
4. python 3.5
5. tensorflow 1.11.0

No need to install

usage: run.py [-h] --net NET [--epochs EPOCHS] [--batch_size BATCH_SIZE] [--gpu GPU] [--img_size IMG_SIZE] [--load_weights LOAD_WEIGHTS]

A testing framework for semantic seg.

optional arguments:
  -h, --help            show this help message and exit
  --net NET             The type of net work which is either unet, deeplab or custom.
  --epochs EPOCHS
  --batch_size BATCH_SIZE
  --gpu GPU             The id of the gpu used when training.
  --img_size IMG_SIZE   The size of input image
  --load_weights LOAD_WEIGHTS
                        Use old weights or not (named net_imgSize.h5)

python run.py --net deeplab --epochs 1000 --gpu 0 --img_size 192

Train a deeplab model with input image size of 192

python run.py --net unet --epochs 1000 --gpu 0 --img_size 192

Train a unet model with input image size of 192

python run.py --net custom --epochs 1000 --gpu 0 --img_size 192

Train a custom model with input image size of 192

• run.py

1. constructs models
2. uses generator to feed images into models
3. use pycocotools to evaluate the result

• basic_model.py

  • unet
  • deeplab

  for run.py to train and evaluate.

• model.py

  • custom model

  When defining your own model, you should implement class custom_model in the model.py which accepts input_size which is the input image size and classes which is the number of classes to be classified. Class custom_model should return a keras model instance so that run.py could build model from it.

• mrcnn/

  • mrcnn

  again for run.py

• config.py

Use a Config class to manage the configuration

• data.py

contains a generator with imgaug for custom model, unet and deeplab and a dataset reader for mrcnn.

• utils.py

contains some utilties that will be used by any file. It contains padding_and_cropping function for inference stage.

Result will be showed on the screen