Real-Time Cascade Template Matching for Object Instance Detection

Abstract

Object instance detection finds where a specific object instance is in an image or a video frame. It is a variation of object detection, but distinguished on two points. First, object detection focused on a category of object, while object instance detection focused on a specific object. For instance, object detection may work to find where toothpaste is in an image, while object instance detection will work on finding and locating a specific brand of toothpaste, such as Colgate toothpaste. Second, object instance detection tasks usually have much fewer (positive) samples in training compared to that of object detection. Therefore, traditional object instance detection methods are mostly based on template matching.

This paper presents a cascade template matching framework for object instance detection. Specially, we propose a three-stage heterogeneous cascade template matching method. The first stage employs dominate orientation template (DOT) for scale and rotation invariant filtering. The second stage is based on local ternary patterns (LTP) to further filter with texture information. The third stage trained a classifier on appearance feature (PCA) to further reduce false-alarms. The cascade template matching (CTM) can provide very low false-alarm-rate comparing to traditional template matching based methods and SIFT matching based methods. We demonstrate the effectiveness of the proposed method on several instance detection tasks on YouTube videos.