Scaled-YOLOv4: Scaling Cross Stage Partial NetworkMotivationYOLOv4 have CSP architectures in backbone, but Neck does't have CSP architectures. Authors want to apply CSP approach, scales both up and down in YOLOv4 for high AP and low inference time. Since the computing facilities rely on restriction like bandwidh, dram of hardware, effective object detector is important.  Main IdeaThe most common..

MotivationThey want to improve YOLO. So They aim three purpose.1. Develope efficient and powerfull model.2. Verify the influence of state-of-the-art Bag-of-Freebies and Bag-of-Specials methods3. Modify state-of-the-art method more effecinet and sutalbe for single GPU training Main IdeaObject Detector StructureA modern detector is usually composed of two parts a backbone and head. Their backbone ..