Custom - Object - Detectiom

Dataset

For the custom images, i have used rubic cube as my custom object. The images are present here.

Traing-Test split

Use this script to split and make train.txt and test.txt. I have put random 10% of data for test.

NOTE : train.txt and test.txt is already provided for my dataset.

Annotation

I have used labelImg to create the desired bounding box, it saves the bounding boxes as xml file. But YOLO model has the following format for bounding boxes

<object class> <x-center> <y-center> <width> <height>

To convert the xml files into desired .txt file, one can use this.

Training

I have used forked version of darknet for the training purpose.

Download the pre-trained model.
Create .data file as :

   classes = 1
   train  = <path to train.txt> 
   valid  = <path to test.txt>
   names = <path to object.names>
   backup = <path to save weights>

Here, Classes : The number of classes we wan to detect(1 in my case). names : names of the class saved as .names

Edit the YOLO configuration file as per your hardware capability.

 a. [net]
    # Testing
    #batch=1
    #subdivisions=1
    #Training
    batch=64  (number of images to be used per batch)
    subdivisions=8 (used to create mini batch : 64//8 in my case)
 
 b. learning_rate=0.001
    burn_in=150 (to control training speed for 150 batches in my case)
    max_batches = 30000 ( maximun number batched to feed in network)
    policy=steps (to control how the leaning rate will decrease)
    steps=3500,16500 (at what steps the learning rate will decrease)
    scales=.1,.1 (with what rate the learning rate will decrease on mentioned steps).
  
 c. Data augmentation :
    angle=0
    saturation = 1.5
    exposure = 1.5
    hue=.1

Use the following command to train the netwrok.

./darknet detector train rubic-cube.data  yolov3-tiny.cfg ./darknet53.conv.74 > train_log.txt

To monitor loss you can use

grep "avg" train_log.tx

To stop the training.

Testing

To test the model use :

./darknet detector test <rubic-cube.data> <yolov3-tiniy-test.cfg> ./<yolov3 weights>

NOTE : I ahve provided my trained weights

Results

You can find more in results directory.