ssd_detector

SSD Object Detection

Train SSD detection model

To train a Single Shot Detector, jump to training_toolbox/ssd_detector directory. You'll see the ssd_detector folder with sample code demonstrating how to train a MobileNetV2-based SSD object detector.

We provide 2 predefined configurations:

• Vehicles and license plates detector.

  • Configuration file: training_toolbox/ssd_detector/vlp/config.py.
  • Trained model: MobileNet v2 0.35 256x256.

• Object detector trained on the COCO dataset.

  • Configuration file: training_toolbox/ssd_detector/coco/config.py.
  • Trained model: MobileNet v2 1.0 256x256.

Quck start with vehicles and license plates detector

The sample model will learn how to detect vehicles and license plates on BitVehicle dataset. The resulting model can further be used to build a sample application, similar to "security barrier" from OpenVINO™.

To train a model, go through the following steps:

1. Download training data and put it in the data/bitvehicle directory according to data/bitvehicle/README.md file. Annotation files in the COCO format (refer to cocodataset for details) are already located in data/bitvehicle.

2. If necessary, you can modify training settings by editing training_toolbox/ssd_detector/vlp/config.py or leave them by default. For more details please read comments in config.py. Notable parameters in train class are:

   • batch_size - number of images in training batch, by default it's set to 32, but could be increased or decreased depending on the amount of available memory.
   • annotation_path - path to json file with annotation in the COCO format, by default it's set to relative path to bitvehicle annotation, but you could use your own annotation.
   • steps - number of training iterations
   • execution.CUDA_VISIBLE_DEVICES - Environment variable to control cuda device used for training. By default, it's set to 0. In case you have multiple GPUs, you can change it to the respective number, or leave this string empty, if you want to train on CPU.
   • cache_type - type of input data to save in cache to speed-up data loading and preprocessing. By default it's set to ENCODED. Remember that caching might cause system slowdown, so if you don't have enough RAM memory better to disable it, pass NONE to this parameter.

3. To start training go to training_toolbox/ssd_detector directory and type in command line:

   python3 train.py vlp/config.py
   

4. To start evaluation process go to training_toolbox/ssd_detector directory and type in command line:

   python3 eval.py vlp/config.py
   

   Do step 4 in another terminal, so training and evaluation are performed simultaneously.

5. Training and evaluation artifacts will be stored by default in training_toolbox/ssd_detector/vlp/model. To visualize training and evaluation, go to training_toolbox/ssd_detector/vlp and run tensorboard with:

   tensorboard --logdir=./model
   

   And view results in a browser: http://localhost:6006.

   

6. When training is complete, model from the checkpoint could be infered on input data by running training_toolbox/ssd_detector/infer.py:

   python3 infer.py vlp/config.py --video --input=<path_to_input_video> --show
   

7. Finally, trained model could be converted to Inference Engine format for optimized inference. To export, go to training_toolbox/ssd_detector and run export.py:

   python3 export.py vlp/config.py <path_to_mo.py>
   

   If OpenVINO™ was installed in a home directory, then <path_to_mo.py> is ~/intel/computer_vision_sdk/deployment_tools/model_optimizer/mo.py.

   As a result, you'll find three new artifacts in training_toolbox/ssd_detector/vlp/model/ie_model:

   • graph.pb - TensorFlow frozen graph,
   • graph.xml and grapn.bin - Inference Engine representation of the model.

8. Model in IR format could be infered using python sample from OpenVINO™ which could be found here: <path_to_computer_vision_sdk>/inference_engine/samples/python_samples/object_detection_demo_ssd_async.py

   python3 object_detection_demo_ssd_async.py -m <path_to_converted_model>/graph.xml -l <path_to_computer_vision_sdk>/deployment_tools/inference_engine/lib/ubuntu_16.04/intel64/libcpu_extension_avx2.so -i <path_to_input_video>