A Study on Object Detection using Faster R-CNN, Retina net, RPN_R_50_FPN, RPN_R_50_C4

The computer vision concept that involves the detection and characterization of objects in images is known as object detection. This research mainly performs this function using four popular models faster R-CNN, RetinaNet, RPN_R_50_FPN, and RPN_R_50_C4 across multiple datasets namely udacity self-driving car dataset, tomatOD dataset, and chess piece detection dataset. Faster R-CNN is famous for its outstanding accuracy although RetinaNet employs a unique focal loss function to balance between speed and accuracy. The purpose of the models RPN_R_50_FPN and RPN_R_50_C4 is to generate region proposals with high quality using different stages in the ResNet-50 backbone. For normalization techniques, the datasets are preprocessed in this research. Instead of training the models from scratch, transfer learning is used whereby starting with weights that were pre-trained on the COCO dataset and after that fine-tuning the required datasets. Accuracy was quantified by evaluating two efficiency measures which are minimum loss and total training duration. Across the different datasets, various performances were demonstrated by different models.