object_detection_kangaroo

Darknet based yolov3 training and Inference in Tensorflow for single class (Kangaroo)

Credits:

I referred and used these fantastic repos during the implementation:

AlexeyAB/darknet

wizyoung/YOLOv3_TensorFlow

Training:

1. Dataset preparation and data augmentation. I have used free Roboflow for format conversion and augmentation

2. Training with 291 training dataset and 28 test dataset.

Training Parameters:

learning_rate=0.001 max_batches = 6000 steps=4000,5400

Inference:

There are two steps:

1. Weight Conversion from darknet weights to Tensorflow checkpoint

2. Testing on sample dataset taken from Kaggle sample_data

#Some results: