Prepare training dataset for Tensorflow

Set up environment

Make sure that all the python dependencies used by the tool are installed for using it.

• Install conda or miniconda (refer to the document on how to install it
• Run conda env create -f environment.yml to create the environment. You can find the environment.yml file in the repo.
• Activate the environment by running source activate tfrecord_dataset

Label Images

• Resize images to unified size. The max width or height of VOC images are 500px. If the image size is too big, it requires larger memory to load these images. You may find it difficult to use a bigger batch size due to memory constraints. Put resize.py in the same folder with the images and run python resize.py. All the images will be resized and saved in resized folder.

• Go to download the labelImg tool from Github

• Label objects in images

• Keep the images in folder named JPEGImages and labels in folder called Annotations. The folder structure should be like below. Replace test with any name you want, such as zombie_train_images etc.

    test_train_images 
    |----JPEGImages
    	    |----0001.jpg
    		 ----0002.jpg
    		 ----...
    		 ----0100.jpg  
    |---Annotations
          |----0001.xml
           ----0002.xml
           ----...
           ----0100.xml  
   test_val_images
   |----JPEGImages
    	    |----0001.jpg
    		 ----0002.jpg
    		 ----...
    		 ----0100.jpg  
    |---Annotations
          |----0001.xml
           ----0002.xml
           ----...
           ----0100.xml  
  

Create tensorflow's record file

Labelled training images and annotations are supposed to be converted to record files. Based on the scripts in this post, I made the script more automated.

• Check out the repo
• Copy test_train_images and test_val_images folder to under the tfrecord_generator folder
• Change the classes in map.txt under tfrecord_generator folder. Remove the default classes and add one class per line. eg. zombie person bigfoot
• Run script ./create_record.sh test . Replace test with the prefix of the training images folder name. For instance, if the training and evaluation folder named "test2_train_images" and "test2_val_images", replace the parameter with test2 instead of test.
• Voila, you will see two record files named test_train.record and test_val.recordrespectively in the same folder

You can find more info on how to enable GPU graphic card and how to train models in these two wiki pages.

• Install tensorflow gpu 1.6 for Ubuntu 16.04
• Train your own model with SSD MobileNet