3D space gaussian clustering question

[smileyenot983]

Is this approach extendable into 3D gaussian-level clustering of gaussian. I mean without identity-alpha bleding and convolutional classifier layer.
For example, if I would like to create a map where key - track_id received from DEVA and values are the ids of gaussians

[smileyenot983]

Upd. I found out that same conv network can be applied to 3d points as the input dimensionality(by default=16) is same.
Issue might be closed

[Caenorst]

I just tried to extract a 3D segment by applying the classifier as a linear layer, but results are not great. Maybe I'm missing something?

classifier = torch.nn.Linear(16, 256, device='cuda')
state_dict = torch.load(classifier_path)
classifier.load_state_dict(OrderedDict(
    weight=state_dict['weight'][:, :, 0, 0],
    bias=state_dict['bias']
))

sel_id = 0

gaussians_logits = classifier(gaussians.get_objects)
gaussians_mask = torch.argmax(gaussians_prob, dim=-1).squeeze() == sel_id