摘要： 自动部件标注是一项复杂的视觉识别任务,但传统训练算法不适用于分布差异下的参数学习。为此,将部件标注描述为基于结构化输出的分类问题,提出一种支持结构化模型的自适应学习算法。通过引入基于相似度的正则算子,重新定义结构化支持向量机的损失函数,使训练损失度和源-目标参数差异度同时最小化。实验结果表明,与传统监督学习算法相比,该算法可使标注准确率提升2%~4%,同时指出部件位置特征的分布差异相比外观特征对自适应学习性能的影响更大。

关键词: 领域自适应, 部件标注, 结构输出, 结构化支持向量机, 目标识别

Abstract: Automatic part annotation is a challenging task in object recognition,but classical training algorithms can not perform well at the parameter optimization when the data distribution in test domain is different from that in training domain.Aiming at this problem,this paper describes the annotation task as structural classification,and proposes a new adaptation algorithm to support the structured output prediction.It re-defines the loss function of Structural Support Vector Machine (SSVM) by introducing a new regularizer based on parameter similarity.The objective is to minimize the optimization loss and simultaneously find the target parameters most similar to source ones.Experimental results show that the proposed algorithm outperforms the traditional SSVM 2~4 percent on average precision.The results also suggest that the difference of pose distribution has more influence on learning performance than the difference of appearance distribution.

Key words: Domain Adaptation(DA), part annotation, structured outputs, Structural Support Vector Machine(SSVM), object recognition

中图分类号: 

• TP18