the accuracy diseases after adding intra-camera loss

[wqydyh]

Sorry to bother you ,after adding intra-camera loss,I find the accuracy diseases whichever combination I use .If possible ,would you please tell me the reason ?Thank you very much.

[wqydyh]

which loss function does the intra-camera loss conflict with

[chenhao2345]

Hi. I use a cluster center loss, which already contains intra-camera information. https://github.com/chenhao2345/ICE/blob/a206eb9a97ad431ab9d9cf38cdcf5ab6fdc6ad1c/ice/trainers.py#L151

[chenhao2345]

So adding the intra-camera loss only makes the model focus more on intra-camera matching.

[wqydyh]

thank you very much for your reply

[wqydyh]

So adding the intra-camera loss only makes the model focus more on intra-camera matching.

sonrry to bother you again ,i use a batchsize of 32 with two 2080 ti Gpu,but my accuracy rate is 2% lower than yours with the same method.Is it the number of graphics cards that causes the effect,？I noticed that you are also using a 32 size batchsize.

[wqydyh]

what's more ,the part of in-tra camera loss has been proved to be useful as well

[chenhao2345]

I think so. Because my conrigurations are tuned on 4 Gpus, which means 8 images per GPU. This setting usually has some affects on results.

[wqydyh]

thanks for your reply, I find that themarket 1501 is probably better suited for four cards, while the MSMT17 is better suited for two. With no IBN，the latter can achieve 39.5 mAP with two cards using baseline plus hard losses.Affter adding 0.1 intra camera loss，the mAP can achieve 40.3%.

[chenhao2345]

Thank you. Glad to know this.