YOLO v2

• This updated version (2016) uses a CNN backbone called Darknet-19

  • a variant of the VGG with simple progressive convolution and pooling layers.

• Anchor boxes to determine the final bboxes.

• Batch normalization to improve accuracy and stability.

• Multi-scale training: training images at multiple scales and average predictions