Object Detection Using Faster R CNN with ROI Alignment using Bilinear Interpolation

Object detection is one of the essential tasks in computer vision with various applications in the realms of surveillance, autonomous driving, robotics, and smart systems. In this paper we detail an object detection system using Faster R-CNN on the COCO dataset. Our system uses a convolutional backbone network to create feature maps, a Region Proposal Network (RPN) to generate regions of interest for candidates, followed by ROI pooling, classification, and bounding box regression to detect and localize the object. The methodology includes relevant algorithms, including Intersection over Union (IoU) for area measurement algorithms, Non-Max Suppression (NMS) for identifying predictions in overlapping areas and for rejecting predictions in those areas, applicable classification and regression loss, for example. Testing on a dataset sampled COCO showed that as a region-based deep learning methodology, Faster R-CNN can achieve the accurate detection of even many specific categories and specific images, while having correspondingly higher computational complexity. The results indicate that object detection using region-based deep learning methodologies is still a viable methodology for accurate object detection, which more importantly describes a number of opportunities for future avenues of improvement and efficiency gains involving optimization, as well as hybrid methodologies. This work also establishes Faster R-CNN as a viable framework to help with both ongoing research and the application of object detection in use.