Investigation of Padding Schemes for Faster R-CNN on Vehicle Detection

Faster Region-based Convolutional Neural Network (R-CNN) is a state-of-the-art object detection algorithm. This method has an excellent performance influenced by several parameters such as number of convolution layers, region proposal algorithm, epoch, and padding used. In this paper, we investigate the impact of padding scheme with AlexNet-based architecture on Faster R-CNN for vehicle detection. The use of padding in the convolution layer prevents reduced spatial information so that optimal features are obtained. We propose same-padding and valid-padding scheme to improve the performance compared with the original AlexNet configuration. The AlexNet model uses a valid-padding scheme at the beginning of the layer and same-padding for the next layer. The performance evaluation of the proposed method is obtained by using our dataset contained with images and videos which represent the original traffic condition in Indonesia. Based on the experiment results, the performance of same-padding scheme has improved by 2.7% compared to the previous model with mAP of 76.5 at the 100th epoch.