Keypoints and part segmentation for inference

[edrdos101]

Hi,

Thanks for providing the source code for your project. I was trying to test this on some unseen data to understand how this works cross-domain and also how it could be used in the wild but I have the following questions:

1. For inference, how do you incorporate keypoints and part segmentation data. Is this even used for inference or just for training?
2. In the paper you mention adapting models to target domain, could you point me to a resource explaining how's this done?

So far I've only tried PCB trained on market-1501 and my objective is to build a custom code that uses your code as an API to track a person in my own dataset across multiple frames

Thank you

[huanghoujing]

Hi, thanks for your attention.

1. For new images, keypoints and part segmentation should be extracted using keypoint estimation and part segmentation models respectively.
2. By adapting a model to target domain, I mean how to effectively use the unlabeled target domain images. PS constraint on target images is the proposed method in this paper.

[edrdos101]

Thanks, just to follow up on that question. Does this mean that adapting to unseen target domain doesn't require any additional steps? If I understood correctly, PS constraint is applied during training, and I can't see how PS constraint can be applied at inference?

Also, could you provide a bit more explanation on what does 'cft_trainer' do?

Thanks!