CenterNet和Pose和Transformer

Transformer综述：

https://mp.weixin.qq.com/s?__biz=MzUxNjcxMjQxNg==&mid=2247514162&idx=2&sn=d094eecbfd91ca1e478c41e29f2b98d5&scene=21#wechat_redirect

https://mp.weixin.qq.com/s?__biz=MzUxNjcxMjQxNg==&mid=2247514982&idx=2&sn=7e38021234b7ab5455429e4485128efd&chksm=f9a1c9e9ced640ff045d1c4fe9d4e98a785602d980b25df4fa18477dd2b4b829ed4fc3fd028f&mpshare=1&scene=23&srcid=0202GgVOuglYZW8ywWAnU2Js&sharer_sharetime=1612267891163&sharer_shareid=1c20b220b96b7bb98298f8f63e40b424#rd

CenterNet

https://mp.weixin.qq.com/s?__biz=MzIwMTE1NjQxMQ==&mid=2247556744&idx=2&sn=07179147834a5a2132f5ca2ac6c11f74&chksm=96f06cdca187e5ca083a86a8cec15d1a85d1a95a2312a1b46706f2a62ac8bab7bcc4b5721d3c&mpshare=1&scene=23&srcid=0202n75x5qRvU11P4unfTZCG&sharer_sharetime=1612267705628&sharer_shareid=1c20b220b96b7bb98298f8f63e40b424#rd

https://mp.weixin.qq.com/s?__biz=MzA4MjY4NTk0NQ==&mid=2247489089&idx=1&sn=e4e9502ff34fbe77180e2e319cde8e5d&chksm=9f80acd7a8f725c173d53b3cf0bfa70d77c12a4279864b4869b832bea376695ecd01caf56d98&mpshare=1&scene=23&srcid=0202a6vB0mZNpj7ut25RSytw&sharer_sharetime=1612267797837&sharer_shareid=1c20b220b96b7bb98298f8f63e40b424#rd

https://blog.csdn.net/weixin_44791964/article/details/107748542

CenterNet与YOLOv3对比实验：

CenterNet相较于YoloV3原版提升比较明显，但是针对改进YoloV3-spp 提升不明显，也低于U版 YoloV3-spp。但是资源占用有优势
https://mp.weixin.qq.com/s?__biz=MzI2OTg2NTI4Mw==&mid=2247483801&idx=1&sn=b8397d4d1935d993180f38be5a5beabc&chksm=ead899b5ddaf10a3354344ff3b79925f76aa83bc5cfb0c7e0b854f00905dffc029093da7201a&mpshare=1&scene=23&srcid=0202H5PhVLQFNysTkUTvGBUS&sharer_sharetime=1612267832500&sharer_shareid=1c20b220b96b7bb98298f8f63e40b424#rd

CenterNetPose：

通过对中心点的偏移来参数化每个关键点。


使用focal loss

We ignore the invisible keypoints by masking the loss.
通过掩盖损失来忽略不可见的关键点。
使用自底向上方法。（associate embedding分组）
将我们的初始预测捕捉到这个热图上最接近的检测关键点。在这里，我们的中心偏移作为一个分组线索，将单个的关键点检测分配给它们最接近的人实例。