CornerNet: Detecting Objects as Paired Keypoints

 https://arxiv.org/abs/1808.01244

 

The corner's network topology is as below

The important thing is the 'Corner Pooling' layer.

At the end of the network, you can make Heatmaps, Enbeddings and Offsets.

Heatmaps will have confidence values of each corner point.

Emeddings will connect each left-top and rifht-bottom corner pair.

Offsets will adjust x and y coordinate values of each corner point.

 

When compared to SSD, cornernet's shape size is larger. I think it could be a trade-off factor with anchors of ssd.

 

You may wonder how it can calculate each corner position. I think it's strongly related to the corner pooling layer. For your reference, even if the corner point is hidden as placed of the edge area. it doesn't matter.

 

 

 

It's the way to make the corner pooling layer. The output layer will be used for heatmaps, embeddings and offsets.