How training data influence the recognition performance?

Abstract

Eff ective fusion of multispectral images captured by visible and infrared cameras enables robust pedestrian detection under various surveillance situations (e.g., daytime and nighttime). However, the performance of detecting small-sized pedestrian instances is still not satisfactory, while small pedestrian detection is important for self-driving and drone vision. Therefore, our effort focuses on improving the detection performance of small-sized multispectral pedestrians which are relatively far from the camera. Since existing multispectral pedestrian datasets mainly consider the large size pedestrians of 50 or more pixels in height, we generate a multispectral pedestrian dataset, named HH (Hanyang and Huins), in which the pedestrian height is from 25 to 50 pixels. To balance the trade-off between the detection performance and speed, we investigate a fusion network to combine two single-shot detectors (SSDs) for the fusion of visible and infrared inputs. The proposed fusion network is trained on public KAIST and KAIST + HH datasets, respectively. From the experimental results, we can observe that the detection performance has been improved a lot by incorporating HH dataset into KAIST dataset for training. The network trained by the original KAIST dataset has only 7.40% average precision (AP). However, the results can be significantly improved to 88.32% by using KAIST + HH for training. This indicates that training images have a great impact on detection performance. © 2019, IJSTR.