This project is an implementation of TensorFlow Object Detection API for detection of women's apparel. The following six classes are handled in this project:

blouse, dress, hat, pants, shorts, skirt

The number of classes can be extended by configuring the model and training it with more data.

1. Download and install Python 3.6 from https://www.python.org/downloads/release/python-360/

   Create environment variables PYTHONPATH, and add the path of directory Python36; Python36\Lib; Python36\DLLs; Python36\Scripts to PYTHONPATH. Add %PYTHONPATH% to PATH environment variable.

2. Install TensorFlow

   pip3 install --upgrade tensorflow

3. Install other dependencies

   pip install Cython
   pip install pillow
   pip install lxml
   pip install jupyter
   pip install matplotlib

4. Download Google Protobuf Windows v3.4.0 release “protoc-3.4.0-win32.zip” from https://github.com/google/protobuf/releases

   Extract the Protobuf zip file to Program Files folder, i.e., C:\Program Files\protoc-3.4.0-win32

5. Clone the TensorFlow model repository.

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

Add the path of models\research; models\research\slim; models\research\object_detection; models\research\object_detection\utils to PYTHONPATH environment variable. Restart computer to update PATH variable.

cd to the path of models\research

  cd path_to_tensorflow\models\research

Execute the following command to compile Protobuf libraries:

   "C:\Program Files\protoc-3.4.0-win32\bin\protoc.exe" object_detection/protos/*.proto --python_out=.

6. Clone this WomenApparelDetection repository.

   git clone https://github.com/jimingh/WomenApparelDetection.git

cd to WomenApparelDetection directory, copy and paste the images that will be segmented into InputImages folder. Execute python imageSeg.py, which will generate segmentation results. Some examples are shown below.

Faster R-CNN, R-FCN and SSD currently are the three best and most widely used object detection models. TensorFlow Object Detection API is utilzed in this project to implement the object detection models. Actually I intended to use Faster R-CNN models at the beginning as Faster R-CNN can achieve the best accuracy in general. Unfortunately, I found there is a bug of the Object Detection API for Python 3 compatibility which does not allow training of Faster R-CNN model faster_rcnn_inception_resnet_v2_atrous_coco in Windows OS. Considering the fact that I do not have a Linux system in hand, I choose SSD model instead, which has faster speed and comparable accuracy.

ssd_mobilenet_v1_coco model has been fine tuned to detect women's apparel in this project. Six classes have been handled, i.e., blouse, dress, hat, pants, shorts, skirt. For each class, 200 images are selected, re-sized and labeled manually to train the model. Some of the training images are from interenet, but the majority are from the product images provided. This is because the women's apparel in the product images is mostly in Asian style, while the images from interenet could be diverse. The model has been trained using Google Cloud ML Engine for around 10 hours, and the loss is smaller than 2.