Generating Pedestrian Training Dataset using DCGAN
Daeun Dana Kim, Muhammad Tanseef Shahid, Yunseong Kim, Won Jun Lee, Hyun Chul Song, Francesco Piccialli, Kwang Nam Choi · 2019
Recently, as autonomous cars are developing very fast, it is the most crucial task to detect pedestrians for autonomous driving. Convolution neural network based on pedestrian detection models has gained enormous success in many applications. However, these models need a large amount of annotated and labeled datasets for training process which requires lots of time and human effort. For training samples, the diversity and quantity of datasets are very important. The proposed framework is based on Deep Convolutional Generative Adversarial Networks (DCGAN), able to generate realistic pedestrians. Experimental results show that DCGAN framework is able to synthesize real pedestrian images with diversity. The synthesized samples can be included in training data to improve the performance of pedestrian detectors. 24,770 images including PETA dataset, Inria dataset were used for the training process.