Object Detection Boosting using Object Attributes in Detect and Describe Framework

摘要

Different objects have unique attributes, visual appearances and physical properties which help human visual system to recognize them better. But can object attributes help improve the object detection performance in computer vision? To answer this very question, we carry out extensive experimentation in this research work and claim that, indeed, object attributes improve the object detection performance significantly. We train feature pyramid networks to learn deep convolutional features for objects and their attributes. When used in combination with each other to infer bounding boxes and class scores for objects, these convolutional features show that object detection boosts significantly. We present a new method to boost the performance of object detection using object attributes in Detect-and-Describe (DaD) framework. We explain multiple approaches for boosting of object detection using their attributes. In these approaches, the convolutional features of attributes are merged with convolutional features of bounding box and class labels using different feature merging techniques to boost object detection. To report the performance of object detection boosting using DaD framework, we train our experimental models on aPascal train split and report performance on aPascal test split. Our results show that object attributes can help boost mean average precision (mAP) of object detection as significant as 2.68%.