基于视觉信息的目标检测和识别模型在训练时往往依赖于来自于训练样本的视角信息,然而附带了视角信息的训练样本通常只有很少的数据库可以提供。当此类信息缺失时,传统的通用目标检测系统通常通过一些非监督学习方法来对样本的视角信息进行粗略估计。本文改进并引入了一种选择性迁移学习方法即TransferBoost方法来解决目标视角信息缺失的问题。TransferBoost方法基于GentleBoost框架实现,通过重新利用其他类别样本中的先验信息来提升当前类别样本的学习质量。当给定一个标定完善的样本集作为源数据库时,TransferBoost同时调整每个样本的权值和每个源任务的权值,实现样本级和任务级的两级知识迁移。这种双层迁移学习更有效地从混合了相关源数据和不相关源数据的数据集中提取了有用的信息。实验结果表明,和直接使用传统的机器学习方法相比较,迁移学习方法所需要的训练样本数大大减少,从而降低了目标检测与识别系统的训练代价,扩展了现有系统的应用范围。

Conventional vision based object detection or recognition models mostly depend on the view information of target examples. However, such attached view information is usually limited in several datasets. When the view information is scare, some generic object detection models try to learn the target by evaluating the view information with unsupervised learning methods. In this paper, a selective transfer learning method, TransferBoost, is improved and introduced to relieve the lack of object view information in training. The proposed TransferBoost, based on the GentleBoost framework, prompts the performance of learning the target by reusing prior knowledge from other object classes. Given a well labeled training set as source task, TransferBoost can transfer the knowledge on both instance level and task level by adjusting the weights of examples and task simultaneously. Such a combination of two levels transfers extracts useful information more effectively from mixed relevant source tasks and irrelevant source tasks. Our experimental results show, that compared to traditional machine learning methods, transfer learning needs much less training examples thus reduces the training cost of object detection or recognition models and extends the applicability of existing models.