Evaluating the Scalability of a Multi-Object Detector Trained with Multiple Datasets

Background Thresholding is a method used to improvise the training of a model with multiple datasets. The primary objective of this paper is to test the scalability of the Background Thresholding approach incorporated with a proposed object detection model for multiple datasets. The proposed research work has taken the Bosch traffic light dataset for traffic light detection, Tsinghua-Tencent 100K Dataset for Traffic sign detection and Udacity annotated driving dataset for car detection. Different algorithms and deep learning frameworks are experimented to provide a detailed analysis as to which model performs the best for the particular task at hand. The proposed research work is concluded with the Faster-RCNN (Regional-CNN) algorithm, which is coupled with Inception V2 to remain as the best for our approach.