vehicle_attributes

Vehicle attributes recognition: type and color

Setup

Prerequisites

• Ubuntu 16.04
• Python 3.6
• TensorFlow 1.13.1
• OpenVINO 2019 R1 with Python API

Installation

1. Create virtual environment

virtualenv venv -p python3 --prompt="(veh_attr)"

2. Activate virtual environment and setup OpenVINO variables

. venv/bin/activate
. /opt/intel/openvino/bin/setupvars.sh

NOTE Good practice is adding . /opt/intel/openvino/bin/setupvars.sh to the end of the venv/bin/activate.

echo ". /opt/intel/openvino/bin/setupvars.sh" >> venv/bin/activate

3. Install the modules

pip3 install -e .
pip3 install -e ../utils

Train Vehicle Attributes model

To train a Vehicle Attributes model, jump to training_toolbox/vehicle_attributes directory. You'll see the folder with a sample code demonstrating how to train a VA model.

We provide predefined configuration for:

• Vehicle type and color recognition.
  • Configuration file: training_toolbox/vehicle_attributes/cars_100/config.py.
  • Trained model: Vehicle Attributes. A model has two output nodes: . type is 4 components: one for each of 4 types - car, van, truck, bus. The resulting type is argmax of these values. . color is 3 components: LAB format normalized from 0 to 1

For demonstration purpose only a small dataset is uploaded cars_100. It is divided by two subsets of images for training and evaluation and serves just as an example of flow. The resulting model is trained on the internal dataset of more than 50000 images of front-facing vehicles in different lighting and weather conditions where number of white/black/gray vehicle colors more than twice exceeds others. It has about 74% car images, 15% truck images, 2% bus images and 9% van images. It was evaluated on BitVehicle dataset:

Metric	Value
Color mean absolute error	9.076
Mean absolute error of l color component	19.75
Mean absolute error of a color component	3.353
Mean absolute error of b color component	4.129
Type accuracy - average	0.958
Type accuracy - car	0.946
Type accuracy - van	0.967
Type accuracy - truck	0.952
Type accuracy - bus	0.966

To train a model, go through the following steps:

Prepare dataset

1. Download training data and extract it in data/cars_100/images folder. There are two files with annotations named cars_100_test.json and cars_100_train.json in the 'data/cars_100' directory. The result structure of the folder should be:

   ./data/cars_100/
   ├── cars_100_test.json
   ├── cars_100_train.json
   └── images
   

2. Please make sure that configuration fila training_toolbox/vehicle_attributes/cars_100/config.py contains correct paths to the annotation files.

3. To use the trained model weights, download it to training_toolbox/vehicle_attributes folder, unpack it and set the corresponding configuration flag 'use_pretrained_weights' to True. Please make sure that the correct path to pretrained model is set in the configuration file.

Train and evaluation

1. To start training go to training_toolbox/vehicle_attributes directory and type in command line:

   python3 tools/train.py cars_100/config.py

2. To start evaluation process go to training_toolbox/vehicle_attributes directory and type in command line: NOTE Run in in parallel in another terminal, so training and evaluation are performed simultaneously.

   python3 eval.py cars_100/config.py

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

   tensorboard --logdir=./model

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

Export to OpenVINO

To run the model via OpenVINO one has to freeze TensorFlow graph and then convert it to OpenVINO Internal Representation (IR) using Model Optimizer:

python3 tools/export.py --data_type FP32 --output_dir <export_path> cars_100/config.py

default export path

• lpr/model/export_<step>/frozen_graph/ - path to frozen graph
• lpr/model/export_<step>/IR/<data_type>/ - path to converted model in IR format

Demo

For the latest checkpoint

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

python3 tools/infer.py cars_100/config.py

For frozen graph

python3 tools/infer.py --model model/export_<step>/frozen_graph/graph.pb.frozen \
    --config cars_100/config.py \
    <image_path>

For Intermediate Representation (IR)

python3 tools/infer_ie.py --model model/export_<step>/IR/FP32/vehicle_attributes.xml \
  --device=CPU \
  --cpu_extension="${INTEL_OPENVINO_DIR}/deployment_tools/inference_engine/lib/intel64/libcpu_extension_avx2.so" \
  <image_path>