基于姿态的判别属性学习及在细粒度识别中的应用

Pose-Based Discriminative-Attributes Learning for Fine-Grained Recognition

姿态变化造成同一对象或同类对象的视觉信息差异巨大,成为计算机视觉中对象识别的一大挑战因素.属性表示重在刻画较高的抽象语义特性,具有应对包括姿态变化的复杂环境变化的鲁棒性,但也给属性学习自身带来了较大难度.如何降低属性学习的难度同时提高属性表示的判别力,成为基于属性表示的识别模型的关键,尤其面临对判别属性要求较高的细粒度识别任务.显式地对姿态建模,在不同姿态下学习能够最大化类别间隔的视觉判别属性,最终作为中间表示用于类别识别.最后,在细粒度公开数据集CUB上验证了所提出的基于姿态的判别属性在细粒度识别任务中的有效性.

Commonly existed various posture of object makes great challenges for object recognition in computer vision lit- erature. Sttribute representation shows robust describable ability with clear semantic meaning invariant to changes of en- vironment factors including posture. However, the inherent description advantages of attributes also result big challenges for itself to learn well worked attribute predictor. Consequently, the key issues in attribute learning are to alleviate the dif- ficuhy of predicting attributes and enhance the discriminant ability at the mean time ,which especially important for fine- grained recognition task. By explicitly modeling the posture states and learning discriminative attribute with respect to dif- ferent postures, describable and discriminative attribute can he built for final category recognition. The proposed pose- based discriminative attribute is verified on publicly available fine-grained dataset CUB with advanced performance.