Play with TensorFlow Object Detection

# create workspace folder
mkdir workspace
cd workspace

# create virtual environment
python3 -m venv cv-venv
source cv-venv/bin/activate
# install tensorflow or tensorflow-gpu
pip install tensorflow-gpu

# get source code
git clone https://github.com/ndaidong/tf-object-detection.git

cd tf-object-detection
pip install -r requirements.txt
./init.py

cd tflib
protoc object_detection/protos/*.proto --python_out=.
export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim
cd ..
python tflib/object_detection/builders/model_builder_test.py

For more info:

• TensorFlow Object Detection - Installation
• Setup a lightweight environment for deep learning

# cd workspace
wget http://www.robots.ox.ac.uk/~vgg/data/vgg_face/vgg_face_dataset.tar.gz
git clone https://github.com/ndaidong/vgg-faces-utils.git
tar -zxvf vgg_face_dataset.tar.gz -C vgg-faces-utils
pip install -r vgg-faces-utils/requirements.txt
cd vgg-faces-utils
./script.py -d vgg_face_dataset/files -o output

Useful links:

• Download and annotate images from VGG Faces dataset.

// coming soon

// coming soon

# cd workspace/tf-object-detection
./make_tfrecord.py -d ../vgg-faces-utils/output -e 100 -o temp/data -l PATH_TO_LABEL_MAP_FILE.pbtxt

Arguments:

• -d, --dir: relative path to folder where we store labels and images
• -e, --extract: how many images we want to extract from the whole set. Default: 100.
• -o, --output: relative path to folder where TFRecord files will be saved into. Default: temp/data
• -l, --labelmap: relative path to label map file. Default configs/label_map.pbtxt
• -r, --ratio: ratio of test set / training set. Default: 0.1 (1 image for test, 9 images for training)

Useful links:

• TensorFlow Object Detection - Preparing Inputs

# cd workspace
wget http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v2_coco_2018_03_29.tar.gz
tar -zxvf ssd_mobilenet_v2_coco_2018_03_29.tar.gz -C tf-object-detection/temp/checkpoints

To find more pretrained models:

• Tensorflow detection model zoo

Depending on the selected checkpoint, we use the diffrent pipeline config.

A long list of sample configs are collected here:

• TensorFlow - Sample config

Just download the appropriate config, then specify the following properties:

Path to pretrained model, e.g temp/checkpoints/ssd_mobilenet_v2_coco_2018_03_29/model.cpkt

Path to train.record, e.g temp/data/train.record

Path to label map, e.g configs/face-detection/label_map.pbtxt

Path to eval.record, e.g temp/data/eval.record

Path to label map, e.g configs/face-detection/label_map.pbtxt

More about pipeline here:

• Configuring the Object Detection Training Pipeline

# cd workspace/tf-object-detection
python tflib/object_detection/train.py --logtostderr --pipeline_config_path=configs/PATH_TO_PIPELINE.config --train_dir=PATH_TO_TRAIN_DIR

PATH_TO_TRAIN_DIR is where we would to save trained model into. It will be created automatically if not yet.

Recommended path: temp/models/MODEL_NAME/MODEL_VERSION/train

# cd workspace/tf-object-detection
python tflib/object_detection/eval.py --logtostderr --pipeline_config_path=PATH_TO_PIPELINE.config --checkpoint_dir=PATH_TO_TRAINING_DIR --eval_dir=PATH_TO_EVALUATION_DIR

PATH_TO_TRAIN_DIR: training model directory that we've specified above.

PATH_TO_EVALUATION_DIR: path to where to store evaluating result. It will be created automatically if not yet.

Recommended path: temp/models/MODEL_NAME/MODEL_VERSION/eval

PATH_TO_PIPELINE_CONFIG

# cd workspace/tf-object-detection
tensorboard --logdir=training:PATH_TO_TRAIN_DIR,test:PATH_TO_EVALUATION_DIR

Useful links:

• TensorFlow Object Detection - Running Locally

# cd workspace/tf-object-detection
python tflib/object_detection/export_inference_graph.py --input_type image_tensor --pipeline_config_path=PATH_TO_PIPELINE.config --trained_checkpoint_prefix=PATH_TO_TRAINING_DIR/model.ckpt-INDEX --output_directory=PATH_TO_OUTPUT_DIR

Useful links:

• Exporting a trained model for inference
• “How to deploy an Object Detection Model with TensorFlow serving” by @KailaGaurav

# cd workspace/tf-object-detection
./predict.py -m PATH_TO_FROZEN_INFERENCE_GRAPH.pb -l PATH_TO_LABEL_MAP_FILE.pbtxt

Arguments:

• -m, --model: relative path to exported .pb graph file
• -l, --labelmap: relative path to label map file. Default configs/label_map.pbtxt
• -f, --file: path to image or video file. Only support .jpg, .png, .mp4, .avi, .mkv
• -c, --cam: index of camera. Use this argument for realtime prediction with specified camera.
• -d, --dir: path to image foldes. Use this argument for predict multi images. Default samples
• -o, --output: relative path to output folder. Go with --dir. Default temp/result

• TensorFlow v1.6, v1.7
• Python v3.6.4, v3.6.5
• Ubuntu 17.10

The MIT License (MIT)