Can the modules for group and pedestrian interaction in the code be used independently?

[Pradur241]

I am currently trying to implement a crowd grouping algorithm, and I was greatly inspired by your work. I would like to ask if the code for row-to-row group interaction can be used alone without the prediction part.

[InhwanBae]

Hi @Pradur241,

Of course! You can extract the group indices without a predictor using line 212 in the model_groupwrapper.py. I have checked that any predictors other than trajectories or directly performing supervised training on the indices with BCEloss work fine. I hope it works well in your work.

https://github.com/InhwanBae/GPGraph/blob/49d86eca04aa72095f949609f1a0fc731a6476e0/model_groupwrapper.py#L211-L212

[Pradur241]

Thanks！

[Pradur241]

@InhwanBae The determination of the value of self.th is a subject of my curiosity. How should this value be adjusted to accommodate pedestrian interaction in various scenarios?

[InhwanBae]

@Pradur241 The value of self.th is a learnable parameter that divides whether two people are in the same group or not, based on the feature distance between them. In easy words, in the case that d_type is Euclidean, it determines whether they are in the same group by thresholding based on the actual distance (e.g., 1 meter). When applied to different data, initialize the thresholding parameter to a suitable and affordable value for custom data, and it will optimized after training.

[Pradur241]

OK，thanks for answering