yolo_processing

A collection of Matlab code written by Sven Bambach, used for processing YOLO detections.

Usage

Generating Training Data

use the create_training_data_from_annotations function to read the annotation_data_XYZ.mat files (which are the output of our in-house coding program) and generate a set of files/folders that can then be fed into YOLO.

This function expects 2 arguments:

• folder with the .mat files
• output folder where it will place YOLO training files

For experiment 15, these .mat annotation files are in a folder on Salk:

/data/aamatuni/obj_loc/training_data/matlab/label_tool_HOME/frame_lists

the structure of the output that create_training_data_from_annotations will produce:

• JPEGImages (folder with all the training images)
• labels (folder with .txt files that list all the ground truth bounding boxes)
• training.txt (a file with a path to each training image, one on each line. these are the images in JPEGImages)

Preparing YOLO

To download and compile the YOLO training code:

$ git clone https://github.com/zehzhang/darknet_for_toy
$ cd darknet_for_toy
$ export PATH=${PATH}:/usr/local/cuda-8.0/bin/
$ make
$ mkdir trained_model_weights
$ wget http://pjreddie.com/media/files/extraction.conv.weights

Training YOLO

Assuming the outputs from create_training_data_from_annotations have beeen created and darknet has been compiled, YOLO training is run inside the darknet_for_toy directory with this command:

$ ./darknet yolo train cfg/yolo.train.cfg path/to/training.txt trained_model_weights extraction.conv.weights 2>&1  | tee training.log

This command will take a number of hours to run. When running it, make sure to use nohup or screen so that the training process isn't shut down when your computer loses internet connection, or disconnects from the GPU server for whatever reason.

The training process will produce a number of model weights files, which you can find in the trained_model_weights folder. The one with the highest number is likely the best one (the number reports how many training iterations for any given set of weights).

Testing training results

To generate bounding boxes on new data, you'll need to pass YOLO a list of paths to new images:

$ ./darknet yolo getboxes cfg/yolo.cfg list_of_test_images.txt output.txt trained_model_weights/yolo_19000.weights

Here list_of_test_images.txt is in the same format as the training.txt file mentioned earlier, just a list of paths to .jpg images, and output.txt is where you want YOLO to dump the inference results. In this command you would replace yolo_19000.weights with whatever the weights file with the highest number was, which will be unique to your particular training.