YOLOv3 Paper Walkthrough: Even Better, But Not That Much

YOLOv3 is an incremental improvement over YOLOv2, primarily featuring architectural changes rather than a new underlying algorithm. The core modification is the introduction of Darknet-53, a deeper backbone network that uses residual blocks and replaces max-pooling with strided convolutions. This version employs three detection heads at different scales to improve the detection of small, medium, and large objects by combining feature maps from different layers. It also shifts to a multi-label classification paradigm, using sigmoid activation functions to allow a single object to be assigned multiple class labels simultaneously.