A hybrid loss balancing algorithm based on gradient equilibrium and sample loss for understanding of road scenes at basic-level

• Published in: Pattern analysis and applications : PAA Vol. 25; no. 4; pp. 1041 - 1053
• Main Authors: Su, Tao, Shi, Ying, Xie, Changjun, Luo, Wenguang, Ye, Hongtao, Xu, Lamei
• Format: Journal Article
• Language: English
• Published: London Springer London 01-11-2022; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Balancing; Computer Science; Datasets; Industrial and Commercial Application; Object recognition; Pattern Recognition; Training; Convolutional neural networks (CNNs); Loss balancing; Gradient equilibrium; Vehicle object detection
• ISSN: 1433-7541; 1433-755X
• DOI: 10.1007/s10044-022-01068-1

Object detection is indispensable to visual environment sensing at basic-level. During that detection there are imbalance between the losses of multi-task and multi-object in state-of-the-art algorithms, resulting in the slowdown of training process and low precision. To address this issue, a hybrid loss balancing (HLB) algorithm combined a loss balancing strategy based on gradient equilibrium (LBGE) with a multi-sample loss balancing (MSLB) strategy is proposed. The LBGE strategy increases the accuracy of the Basic network by 3.38% on VOC dataset and by 4.22% on KITTI dataset by updating the weight of each loss during the training iteration. The MSLB strategy with the optimal super-parameter value 200 can improve the accuracy of the basic network by 2.51% on VOC dataset and by 6.62% on KITTI dataset by assigning larger weights to the proposal regions which are more difficult to train. With both strategies working together, the proposed HLB algorithm improves the accuracy by 3.88% on VOC and by 7.24% on KITTI, enhancing robustness to the cross-domain datasets than single strategy. Moreover, the proposed HLB loss function obtains the highest accuracy at 84.02%, a 2.39% higher than that of original loss and other loss functions on average. In a word, the HLB algorithm with the LBGE and MSLB strategies have better understanding ability of basic-level road scenes than Basic network on VOC and KITTI dataset, and can also accelerate the early training speed of the Basic network.