Corner pooling
Ever thought about using corner pooling to further improve AP?
Hi @zhaopings ! How would the corner pooling improve their idea? The corner pooling layer improves corner localization, but here they are detecting extreme points(top, bottom, left and right), not corners. If you have an idea about how they would work, please tell me, I'm looking for ideas for my faculty degree. Thanks!
It's true that corner pooling can't be used for extreme points prediction, but I tried to use center pooling for center point prediction and achieved some performance improvement
------------------ 原始邮件 ------------------ 发件人: "biras11"<[email protected]>; 发送时间: 2020年4月24日(星期五) 凌晨1:38 收件人: "xingyizhou/ExtremeNet"<[email protected]>; 抄送: "Parallel world"<[email protected]>;"Mention"<[email protected]>; 主题: Re: [xingyizhou/ExtremeNet] Corner pooling (#40)
Hi @zhaopings ! How would the corner pooling improve their idea? The corner pooling layer improves corner localization, but here they are detecting extreme points(top, bottom, left and right), not corners. If you have an idea about how they would work, please tell me, I'm looking for ideas for my faculty degree. Thanks!
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.
Thanks for the reply @zhaopings ! I will try to implement center pooling in extreme net too. Did you notice a high increase in processing time when you added the center pooling?
I didn't care much about FPS, but in COCO validation, the speed is not much slower than before
------------------ 原始邮件 ------------------ 发件人: "biras11"<[email protected]>; 发送时间: 2020年4月28日(星期二) 凌晨1:24 收件人: "xingyizhou/ExtremeNet"<[email protected]>; 抄送: "Parallel world"<[email protected]>;"Mention"<[email protected]>; 主题: Re: [xingyizhou/ExtremeNet] Corner pooling (#40)
Thanks for the reply @zhaopings ! I will try to implement center pooling in extreme net too. Did you notice a high increase in processing time when you added the center pooling?
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.