A Transfer Learning Approach to Recognize Pedestrian Attributes
S M Nazmuz Sakib, Anik Sen, Kaushik Deb · 2023
Pedestrian attribute recognition has recently created a significant impact because of its soft bio-metric property to recognize individuals. The popularity of the pedestrian attribute recognition task is gained after deep learning, especially the convolutional neural network. This paper represents a transfer learning approach for the pedestrian attributes recognition task because of its high performance and low training cost. The Mask RCNN object detector extracts the images of isolated 146 pedestrians. After that, the preprocessed images are passed to different CNN architectures, i.e., Inception ResNet v2, Xception, ResNet 101 v2, to extract the spatial features. Experiments reveal that the Xception architecture outperforms the competition with a 90.33\% accuracy rate. Following that, some experiments on the Xception architecture are carried out by freezing the last 4, 8, 12, 16, 20, all, and none layers (excluding the fully connected layers). For the RAP v2 dataset, experimental results show that freezing the last 16 layers provides the best accuracy, 92.52%, outperforming existing methods in terms of accuracy.