search

DoubtedSteam / MPANet

The official implementation for Discover Cross-Modality Nuances for Visible-Infrared Person Re-Identification (CVPR21)
MIT License
59 stars 5 forks source link

How about replacing the CC loss with triplet loss? #3

Open mouxingyang opened 3 years ago

mouxingyang commented 3 years ago

Hi~ It's a nice job. The proposed CC loss seems to narrow the modality gap, which is often done by triplet loss. So, how about replacing CC with the triplet loss?

Thanks~

swb-upc commented 3 years ago

Hi~ It's a nice job. The proposed CC loss seems to narrow the modality gap, which is often done by triplet loss. So, how about replacing CC with the triplet loss?

Thanks~

I compared the two loss functions. CC loss is better than triplet loss In terms of performance.

swb-upc commented 3 years ago

Can you give a pre-trained model for comparison? My training did not achieve the results in the paper.

mouxingyang commented 3 years ago

Hi~ It's a nice job. The proposed CC loss seems to narrow the modality gap, which is often done by triplet loss. So, how about replacing CC with the triplet loss? Thanks~

I compared the two loss functions. CC loss is better than triplet loss In terms of performance.

Did you record the ablation results between them? Could you please give me the results? Thanks~

mouxingyang commented 3 years ago

In my implementation, MPANet with cc loss exceeds that with triplet loss by nearly 10% in terms of the performance. I really wonder why the cc loss could help the performance improvement. Thanks~~

swb-upc commented 3 years ago

加个QQ讨论一下吧。1092654533

邵文斌

@. | 签名由网易邮箱大师定制 On 7/27/2021 @.> wrote：

In my implementation, MPANet with cc loss exceeds that with triplet loss by nearly 10% in terms of the performance. I really wonder why the cc loss could help the performance improvement. Thanks~~

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.