YOLOv3 emerged as an incremental improvement over YOLOv2, introducing a new Darknet-53 backbone and multi-head detection architecture. The authors replaced maxpooling with strided convolutions to preserve spatial information and added residual blocks inspired by ResNet. These changes aimed to address limitations of previous versions while maintaining YOLO's real-time detection capabilities.

Unlike YOLOv2's single detection head, YOLOv3 employs three prediction tensors at different scales (13×13, 26×26, and 52×52) to detect objects of varying sizes. Each grid cell predicts three bounding boxes across 80 classes, generating up to 10,647 possible detections per image. This multi-scale approach significantly improves small object detection while maintaining performance on larger objects.

The architecture also incorporates feature pyramid connections, combining semantic information from deeper layers with spatial details from shallower layers through upsampling and concatenation. YOLOv3 shifts from multiclass to multilabel classification, allowing objects to belong to multiple categories simultaneously. While these modifications improved detection accuracy, the authors acknowledged the changes were relatively modest compared to the dramatic improvements seen between YOLOv1 and YOLOv2.