Analysis of Transfer Learning on Faster R-CNN for Vehicle Detection

Abstrak

Abstract—Computer vision is one of the favorite research topics recently, especially for object detection task. Faster Regionbased Convolutional Neural Network (R-CNN) is a state-ofthe-art object detection algorithm. This method has an excellent performance influenced by several parameters such as the number of convolution layers, epoch, padding scheme, network initialization, etc. In this paper, we perform an analysis of the impact of transfer learning using pre-trained AlexNet on Faster R-CNN for vehicle detection. Transfer learning method enables us to use a small amount of training data and training time to achieve good performance. Based on the experimental results, the performance of transfer learning has signiï¬cant improvement by 15.9% compared to the full-training model with mAP of 73.1% at 10th epoch. Keywords–Convolutional neural network; Transfer learning; Object detection; Vehicle detection