problem of crowhuman datase?

[angryhen]

The crowdhuman dataset doesn't have any ID tag, so how it's training successfully?

[ifzhang]

We give each bbox a unique id and the total ids are larger than 1,000,000.

[zengjie617789]

Can you share the details about training of the self-supervised learning on crowdhuman? thank you

[richardvogg]

I would also be interested in learning more about the self-supervised learning part (especially about the implementation).

Great work and great model!

[zengjie617789]

As far as i am concerned， the self-supervised learning in FariMOT is treating the process of refining object's features as a classification task, that is mean the num of classes is the same as the objects in datasets, such as 100000 in Crowdhuman dataset.

[richardvogg]

Thanks for the fast comment! :) You are probably right. When I read the paper, I was hoping that there would be some more contrastive-learning-type action happening. When I read the quote from the FairMOT paper which I pasted below, I thought there would be several instances of each image, created by random transformations, which could then be used to learn the representation.

"Inspired by [51], we regard each object instance in the dataset as a separate class and different transformations of the same object as instances in the same class. The adopted transformations include HSV augmentation, rotation, scaling, translation and shearing." (51 is this paper)

[GandalfTGrey]

So during the pre-training on CrowdHuman or other image level datasets such as COCO, the "self-supervise" means that only the Re-ID head is self-supervised training, the rest three heads is "supervised training". Did I get it right?

[GandalfTGrey]

And in this designed self-supervised training task, the Re_ID head is encouraged to distinguish as many instances as possible?