DLPNet: Dynamic Loss Parameter Network using Reinforcement Learning for Aerial Imagery Detection

Object Detection on aerial imagery is a challenging task due to the following unique characteristics of the aerial imagery data: the large image size and the massive volume of data. Since the original image size from remote sensing sources is generally more colossal than the typical natural images, training detection models with a small mini-batch size are expected for the many practical, real-world applications. Furthermore, using a small mini-batch size enable the model to have better generalization performance. However, it causes an unstable learning process due to high gradient noise. In this case, reducing the learning rate enable the model to train reliably. However, a small learning rate can cause slow learning and convergence to the local minimum point also. Therefore, in this work, we propose a novel method, DLPNet, to enable robust and stable training with a small mini-batch size and various learning rates. Our model composes of an object detection model and a reinforcement learning agent. Our reinforcement learning agent extracts features from training mini-batch data and determines optimal parameters to the loss function of the object detection model. This dynamic loss function with adaptive parameters can achieve a more robust and stable learning process than the original baseline model. We demonstrate the effectiveness of our approach with a challenging object detection dataset, DOTA-v2.0. In addition, we release our code for reproducibility and to promote further research in this area1.