Partial half fine-tuning for object detection with unmanned aerial vehicles

Abstract

Deep learning has shown outstanding performance in object detection tasks with unmanned aerial vehicles (UAVs), which involve the fine-tuning technique to improve performance by transferring features from pre-trained models to specific tasks. However, despite the immense popularity of fine-tuning, no works focused on to study of the precise fine-tuning effects of object detection tasks with UAVs. In this research, we conduct an experimental analysis of each existing fine-tuning strategy to answer which is the best procedure for transferring features with fine-tuning techniques. We also proposed a partial half fine-tuning strategy which we divided into two techniques: first half fine-tuning (first half F-T) and final half fine-tuning (Final half F-T). We use the VisDrone dataset for the training and validation process. Here we show that the partial half fine-tuning: Final half F-T can outperform other fine-tuning techniques and are also better than one of the state-of-the-art methods by a difference of 19.7% from the best results of previous studies.