Human Interaction-Free Object Localization in a Scene

Object detection methods use NMS (Non-Maximum Suppression) to remove multiple detections for a particular object for its localization. To perform this task, NMS requires a confidence threshold and an IoU (Intersection-over-Union) threshold which need to be supplied by an user. Thresholds are fixed and different for different object detection methods, e.g., R-CNN, Faster R-CNN, YOLO, etc. In this paper, we propose a method that uses a suitable regression model to find the threshold values which is adaptive in nature, eliminating the need for human interaction for localization of objects in the scene. The order of the model is determined through bias-variance trade-off and its goodness-of-fit is justified by R2 (R-squared) score and χ2 (Chi-squared) test. Results are impressive and attractive.