try CornerNet

[tkng]

CornerNet is a recently proposed neural network for object detection.

http://openaccess.thecvf.com/content_ECCV_2018/html/Hei_Law_CornerNet_Detecting_Objects_ECCV_2018_paper.html

The idea of CornerNet looks very interesting. They don't use anchor boxes. They use 2 heatmaps to predict "top-left corner" and "bottom-right" corner. Though it looks a little bit difficult to implement on Blueoil, since it requires a novel operator "corner pooling", but at least it would be worth trying.

• [ ] implement corner pooling on lmnet
• [ ] implement hour glass network (backbone network)
• [ ] implement extraction network
• [ ] implement loss function of CornerNet on lmnet
• [ ] compare result vs YOLOv2 on Pascal VOC
• [ ] compare result vs YOLOv2 on MS COCO
• [ ] implement slimmer version of CornetNet
• [ ] implement corner pooling on dlk
• [ ] implement post processing

This task would be very interesting, but unfortunately I don't enough time to implement...

[lm-konda]

I found example of simple corner pooling implementation (but using max_pooling2d) https://github.com/tensorlayer/tensorlayer/issues/781

[tkng]

This one looks not correct for me...

Please see the Figure 3. of the original paper.

[lm-konda]

Sorry. I will confirm it.

[tkng]

No worry, this issue is not high priotity for now.