Investigation Learning Rate Parameter for Keypoint-Based UAV Object Detection

Learning rate is one of the challenging hyper-parameters during training deep learning model. An optimal model requires proper configuration value of learning rate. Larger learning rate results in rapid changes and makes the model not optimal, whereas smaller learning rate makes the model train slower. Finding a good choice of learning rate can be essential to getting the network to learn at all. In this paper, we proposed to compare the CenterNet model with different backbones and different learning rates. Backbones that we use are ResNet-101 and Deep Layer Aggregation, moreover there are eight experiments in total using different learning rate. CenterNet is a model based on keypoint estimation. In other words, CenterNet tries to find the center point of the object which is then regressed into other object properties such us weight and height of bounding boxes. Furthermore, we focus on detecting aerial images, hence the dataset we use is the VisDrone-2019 dataset. Our result shows that training Deep Layer Aggregation backbone using learning rate of 0.000125 can achieve the highest performances around 18.2 percent.