Hardness Prediction for Object Detection Inspired by Human Vision

Abstract

We introduce eye tracking features including existing features like (1) scan path and (2) heat map, and novel features including (1) components of scan path and (2) peaks of heat to better define and understand human vision. In this paper, these features are used to describe the eye movements of a person when he/she is watching an image and looking for the target object in it. Based on these features, a new image complexity called eye tracking complexity is defined. Eye tracking complexity can be computed either by carrying out eye tracking experiments and extracting eye tracking features or through a convolutional neural network (CNN), which is introduced in this paper. This CNN computes eye tracking complexity directly from images. It has been validated that eye tracking complexity of an image corresponds to the detection algorithms average precision over an image. Thus, eye tracking complexity can be used to predict the hardness of object detection, which can yield guidelines for the hierarchical algorithms design.