Compared with RGB based human detection, the thermal or depth images in the identity-preserved human detection tasks not only contain less information but also with large amount of noise. Besides, the labels in the tasks are weak, that is, plenty of ground truth boxes were inaccurate for learning process. The main contributions lie in three aspects: (1) Dynamic noise fixing Method is proposed to reduce the noise pixels in the input images. (2) In order to deal with the weakly labeled ground-truth boxes, noise robust hard example mining is proposed by combining noise example filtering and online hard example mining together. (3) To tackle multi-source data fusion in the object detection task, we propose three fusion strategies named early fusion, intermediate fusion and late fusion, which could be regarded as image level fusion, feature level fusion and box level fusion respectively. 

Our results of single model with single-scale testing obtained the 2nd place in the Depth and Thermal Fusion track and 3rd place in the Thermal track. More details with better results could be found in the following paper at the competition workshop in FG2020. 

 

Related paper:

[1] Zijian Zhao, Jie Zhang, Shiguang Shan. Noise Robust Hard Example Mining for Human Detection with Efficient Depth-Thermal Fusion. IEEE International Conference on Automatic Face and Gesture Recognition Workshops (FGW), 2020.