Reason for choosing the maxpool (size = 2, stride=1) in yolov3-tiny?

[Sudhakar17]

@AlexeyAB : Max Pool is for reducing the feature map size but if we choose stride=1, it retains the same size. I am talking about the final maxpool layer of yolov3-tiny. Is there any purpose for this operation?

https://github.com/AlexeyAB/darknet/blob/master/cfg/yolov3-tiny.cfg#L93

[AlexeyAB]

• stride=2 in any layer (maxpool, convolutional, reorg, ...) is for reducing the feature map size
• [maxpool] layer is for reducing the spatial dependence of features (it doesn’t matter which of the 4 cells contains the maximum feature) - this allows you to detect compressed / stretched / elastic / slightly rotated objects

You are about - it gives +~2-4 mAP points: https://github.com/AlexeyAB/darknet/blob/ba6c35ac8965cd70804020fb4810133b3d32a5ad/cfg/yolov3-tiny.cfg#L93-L95

This is a smaller version of SPP-block: You can read about Spatial Pyramid Pooling: https://arxiv.org/abs/1406.4729v4

https://github.com/AlexeyAB/darknet/blob/ba6c35ac8965cd70804020fb4810133b3d32a5ad/cfg/yolov3-spp.cfg#L575-L597