• Linux Machine
• cudnn 7.0+
• cuda 9.0+
• tensorflow 1.5+

• Search for the objects which you intend to classify on Google
• Open the console & copy the contents of download_images.js in it. This will download a urls.txt file.
• Run the download_data.py.

mkdir images
python3 --urls urls.txt --output images

• Now you need to annotate the images you just downloaded in the Pascal VOC format.
• Install LabelImg

git clone https://github.com/tzutalin/labelImg.git
cd ~/labelImg

# for python3
sudo apt-get install pyqt5-dev-tools
sudo pip3 install lxml
make qt5py3
python3 labelImg.py

• This would open up a window like below. Press 'w' to draw a rectangle around the area of interest and enter the label. Save the file, by pressing 'Ctrl+s', this would create the corresponding XML file in Pascal VOC format.
• Move all the XML files to the annotations directory.

• We need to convert the XML files to CSV so that it can then be converted to TFRecord files.
• Edit xml_to_csv.py to include the desired csv file name & location

python3 xml_to_csv.py

• Create a directory data & move the mobile_labels.csv to it.

• Here, we create 2 seperate csvs train.csv & test.csv in data folder using the mobile_labels.csv, generated above.

python3 splitting_labels.py

git clone https://github.com/tensorflow/models.git 

sudo apt-get install protobuf-compiler python-pil python-lxml jupyter matplotlib
sudo apt-get install -y protobuf-compiler
cd ~/models/research
protoc object_detection/protos/*.proto --python_out=.

cd ~/models/research
python3 object_detection/builders/model_builder_test.py

If it runs fine, then it will display an OK message. However, if you get the error No module named 'object_detection', then do the following.

cd ~/models/research/object_detection/slim
python3 setup.py build
python3 setup.py install
sudo pip3 install -e .
cd ..
export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim  

• Tensorflow Object Detection API requires that the training data be in TFRecord format.

python3 generate_tfrecord.py --csv_input=data/train_labels.csv --output_path=data/train.record
python3 generate_tfrecord.py --csv_input=data/test_labels.csv --output_path=data/test.record

Downloading the ssd_mobilenet_v1_coco_2017_11_17.tar.gz

cd ~/Object_Detection
wget http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_coco_2017_11_17.tar.gz
tar -xvzf ssd_mobilenet_v1_coco_2017_11_17.tar.gz

Tensorflow provides configuration file for a number of pretrained models. You can check them as

ls ~/models/research/object_detection/samples/configs

Copy the configuration file for ssd_mobilenet_v1_coco_2017_11_17 to our project directory

cp ~/github/models/research/object_detection/samples/configs/ssd_mobilenet_v1_coco.config  ~/Object_Detection/training/ssd_mobilenet_v1_mobile.config

Open the copied file in an editor. Edit the following lines as per the use case.

• Line 9 => num_classes : x (x is the number of labels in step 2)
• Line 156 => fine_tune_checkpoint : "ssd_mobilenet_v1_coco_2017_11_17/model.ckpt" (path where to store the model checkpoint)
• Line 175 => input_path : "data/train.record" (path to training data which is stored in TFRecord format)
• Line 177 => label_map_path: "training/object_detection.pbtxt" (path to pbtxt file, which assigns an integer to each label)
• Line 189 => input_path: "data/test.record" (path to validation data which is stored in TFRecord format)
• Line 191 => label_map_path: "training/object_detection.pbtxt" (path to pbtxt file, which assigns an integer to each label)

• Tensorflow requires us to map a class to a label. So, inside training directory, create a file object-detection.pbtxt and add the following content in it:

item {
  id: 1
  name: 'mobile'
}

python3 ~/models/research/object_detection/train.py --logtostderr --train_dir=training/ --pipeline_config_path=training/ssd_mobilenet_v1_mobile.config

• --train_dir is the directory where your trained model checkpoints will be saved.
• --pipeline_config_path is for the model used for training.

 python3 ~/models/research/object_detection/export_inference_graph.py --input_type image_tensor --pipeline_config_path training/ssd_mobilenet_v1_mobile.config --trained_checkpoint_prefix training/model.ckpt-3517 --output_directory mobile_inference_graph

• --output_directory will then contain the frozen trained model

• Create a directory test_images & add some testing images as image1.jpg, image2.jpg, .... imageN.jpg
• Edit line 53 in mobile_detection.py as

TEST_IMAGE_PATHS = [ os.path.join(PATH_TO_TEST_IMAGES_DIR, 'image{}.jpg'.format(i)) for i in range(1, 5) ]

• Run the mobile_detection.py script, to check the result

python3 mobile_detection.py