Parameter Investigation in Low Computing Cost Model-Based EfficientDet for UAV Object Detection

Unmanned Aerial Vehicles (UAV) are equipped with high-resolution cameras. Previous researches novel object detection models for UAV images, but available models took high computational cost to retrain. Models that not trained using the UAV image will result in poor performance. EfficientDet object detection model is significantly lower in computing cost. In this paper, the author wants to investigate epochs and optimisation function impact to the performance of the low computing cost model-based EfficientDet on UAV images. An object detection system on UAV imagery is trained using conventional computer will be designed to detect 10 classes object in Visdrone dataset using the D0 version of the EfficientDet model. After the data is obtained then preprocessing will be carried out in the form of annotation conversion. Next, model training process will be trained from 10 to 50 epochs. In each training, the model will be tested so as to produce a validation value. The last validation value will be analysed as performance benchmark. As a result, our proposed model surpassed state-of-the-art models in Average Recall score with 2.1 % ARmax1 for Visdrone by only using 50 epochs or less.