chengdazhi / Deformable-Convolution-V2-PyTorch

Deformable ConvNets V2 (DCNv2) in PyTorch
MIT License
1.44k stars 229 forks source link

Questions about offsets #69

Open wWHWw opened 3 years ago

wWHWw commented 3 years ago

Hi ! Excellent work on making this implementation of Deformable Convolution.

When I used DCN to replace some of my backbone conv to train an network for object tracking problem, I found that simply replaced the conv to DCN didn't improve performance as expect and it decreased for about 6-8% on both precession and succese.
When I check the offset value during training, offset mean can reach about 100 or even larger.And the responds feature were much more ambiguous than normal convolution.

So is there any solutions to fix or contain the offsets during learning, so the receptive field can be limited to a fixed area?

YiRuChenbilibili commented 2 years ago

你好,在可变形卷积的使用上我遇到了同样的问题。请问你最后怎么修改的?或者只是单纯地不使用了?

wWHWw commented 2 years ago

你好,在可变形卷积的使用上我遇到了同样的问题。请问你最后怎么修改的?或者只是单纯地不使用了?

不太记得了,印象里是没查出为什么

YiRuChenbilibili commented 2 years ago

好的,谢谢你

@.***

@.*** |

---- 回复的原邮件 ---- | 发件人 | @.> | | 日期 | 2022年04月12日 15:43 | | 收件人 | @.> | | 抄送至 | @.**@.> | | 主题 | Re: [chengdazhi/Deformable-Convolution-V2-PyTorch] Questions about offsets (#69) |

你好,在可变形卷积的使用上我遇到了同样的问题。请问你最后怎么修改的?或者只是单纯地不使用了?

不太记得了,印象里是没查出为什么

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>