Abstract: Vehicle appearance segmentation is a major application of computer vision in the traffic scenes. In recent years, due to the fast development of deep learning technique, CNN based image segmentation methods have obtained great breakthrough. In order to solve the data imbalance problem, this paper proposes a new adaptive margin loss function to replace softmax loss. A CNN model is trained to achieve end-to-end pixel level semantic segmentation. Besides, a vehicle appearance dataset is built for model training and testing. Experimental results demonstrate that the new method shows superior performance, especially for categories with small area.

Key words: adaptive margin loss, convolutional neural network, semantic segmentation, vehicle appearance dataset

摘要： 车辆外观分割是计算机视觉在交通场景中的一个重要应用。得益于近几年深度学习技术的火热兴起，以深度卷积神经网络为主的分割方法在诸多领域内取得了突破进展。针对车辆外观分割中的样本不均衡问题，提出一种新的自适应损失函数，替换了原始的softmax损失，并且训练一个新的卷积神经网络模型，实现了端到端的像素级语义分割。同时构建了一个车辆外观分割数据集，用以模型的训练和测试。实验结果表明，该网络对比同类算法拥有较高的分割准确率，对于面积较小的类别有更好的效果。

关键词: 自适应边距损失, 卷积神经网络, 语义分割, 车辆外观数据集