Traditional detectors like YOLO and Faster RCNN suffer from NMS overhead, occlusion handling issues, scale‑variation challenges, and consequently low recall. I decided to explore RT‑DETR as a real‑time, end‑to‑end transformer‑based alternative that eliminates NMS and promises better handling of occlusions and scale variations.
I often refer back to a paper that remains one of my favourites for building generic object detectors.
Still one of my favourite papers in computer vision for building generic object detectors!