Car color identification

This project is about identification the colors of cars. It is based on CNN architecture. It is implemented with using python 3.7.

Python scripts

1. dataset_builder.py parses XML annotations and creates images in specific dirs according ImageDataGenerator and trainDataGen.flow_from_directory(...). More precisely, it cuts off sub images according bounding boxes from XML annotations and puts them in dirs, matching to few colors.
2. color_net.py contains the CNN architecture and the main functions to train net and predict on images.
3. test_net.py loads weights color_weights.hdf5from this (or trains net) ant tests it making videos from original images and predicted labels.
4. image_checker.py is intended for checking annotations - is it bad or good. It helps to detect wrong labels on the image.

Dependencies

1. Keras 2.0.0
2. Tensorflow 1.14.0
3. OpenCV 4.1.0
4. Numpy 1.16.2

Data sets

1. 109 Коридор мойки.zip
2. 102 Въезд из мойки в ремзону.zip

Problems

1. Some colors (green, brown and grey) are missed at both data sets.
2. Some annotations are distorted: bounding boxes of not cars are labeled as cars, for example: images vlcsnap-2019-08-07-22h56m06s489.png and vlcsnap-2019-08-07-23h09m17s466.png contain bad annotations. To check this just run script image_checker.py.
3. Versions of python packages above don't allow me to save neural network to use it again. Apparently saving functions from model.save() and ModelCheckpoint() don't work. See many related issues here:
   • alexander-rakhlin/CNN-for-Sentence-Classification-in-Keras#13
   • keras-team/keras#4904

So, to see how net works use script test_net.py: it loads color_weights.hdf5 via loadWeights() (or trains the net) and then test it, producing videos with predicted colors on bounding boxes.