Instruction

Setup

• If use GPU, export GPU_ENABLE=1
• script/docker/setup.sh
• Install VNC viewer

Start

• script/docker/start.sh
• docker exec -it demoimageclassification_main bash
• Open VNC viewer with address 0.0.0.0:5900 to see camera

Training

• Run python train_model.py
  • options:
    • --dataset: path to input dataset, default: dataset
    • --model: training model (letnet or minivggnet), default: minivggnet
    • --output: path to output model, default: output/minivggnet.h5
    • --reset: value: 1 - capture images then train, value: 0 - train with current dataset
• Capture pictures:
  • Press SPACE to capture pictures for current class (in camera window)
  • Press SHIFT to move to next class (in camera window)
  • Press ENTER to start training (in camera window)
  • Press Esc to quit (in camera window)
• Tips:
  • Take at least 100 images per class
  • With images number per class less than 1000, I prefer model lenet
  • With images number per class less than 1000, I prefer model minivggnet

Tesing

• Run python test_model.py
  • options:
    • --dataset: path to input dataset, default: dataset
    • --model: training model (letnet or minivggnet), default: minivggnet
• Point camera to object
• Dected Image window will show object with highest match score
• Press Esc to quit (in camera window)