A Resampled Feature Pyramid for Fast Deformable Object Detection

Usui Yutaka, Kondo Katsuya

doi:10.2299/jsp.18.49

In this paper, a fast approximate feature pyramid construction method for deformable part models (DPMs) is proposed. DPMs can describe significant variations in the appearance of objects. Therefore, they are widely used for object detection due to their high accuracy and efficiency. However, their high computational cost is a major issue to be tackled for time-critical tasks, such as the real-time image recognition of video. For such tasks, a fast detection process is desired. In the proposed method, we focus on the optimization of the feature pyramid constructor and replace a conventional constructor with a resampling-based method. The conventional feature pyramid constructor requires multiple recursive rescaling processes, which are very time-consuming. By replacing these rescaling processes with an image-resampling process, we can improve the speed of the object detector. Approximate sampling is effective because the original scaling method requires a moderate image-size reduction. To avoid performance loss, we also propose a feature pyramid constructor framework that combines approximate scaling and conventional scaling. The evaluation results show that the proposed resampling-based feature pyramid method is 7% faster than the conventional method. In addition, this method can be combined with other optimization methods for searching the feature space.

A Resampled Feature Pyramid for Fast Deformable Object Detection

Search this article

Description

Journal

References(5)*help

Keywords

Details 詳細情報について

Export

Report a problem