human-3d / Human3D

We propose the first multi-human body-part segmentation model, called Human3D 🧑‍🤝‍🧑, that directly operates on 3D scenes. In an extensive analysis, we validate the benefits of training on synthetic data on multiple baselines and tasks.
https://human-3d.github.io/
54 stars 6 forks source link

Preprocessing Problems #17

Open jako5 opened 4 months ago

jako5 commented 4 months ago

Hi everybody, there seem to be multiple issues going on, which are unfortunately preventing me from doing any tests with the repo.

  1. The current specification for preprocessing the "synthetic_humans" dataset is asking me to pass
    --dataset="synthetic_humans" 

... as an argument, while the preprocessing script itself

        assert self.dataset in [
            "egobody",
            "synthetic",
        ], f"{self.dataset} is not supported."

...does not allow this argument.

Furthermore, the current version of the test_set_release dataset does not seem to contain any split_train.txt (or train_list.txt) files that are required to split the datasets during preprocessing. I really hope you can push the required updates, as it would be great to do some tests with this :)

jako5 commented 4 months ago

From what I can gather, for the current codebase to work with the current dataset, the file structure has to look like this for the synthetic dataset:

image

and the splits beeing extracted from the 60GB dataset, renamed and placed in the data folder like this

image