how was octree generated?

[junmeng6025]

Hi, thanks for the brilliant work first!

I'm a bit confused about how was octree generated. In classification.py it seems that octree already exists in the batch dictionary,

def forward(self, batch):
    octree, label = batch['octree'].cuda(), batch['label'].cuda()
    data = self.get_input_feature(octree)
    logits = self.model(data, octree, octree.depth)
    log_softmax = F.log_softmax(logits, dim=1)
    loss = F.nll_loss(log_softmax, label)
    pred = torch.argmax(logits, dim=1)
    accu = pred.eq(label).float().mean()
    return loss, accu

which looks like something from dataset.

But logically octree is generated upon the input point cloud, rather than being provided along with dataset. Could you explain the origin outcoming of octree, i.e. when/in which script was it generated and written into batch dict? Thanks a lot!

[Aristo23333]

Hi, thanks for the brilliant work first!

I'm a bit confused about how was octree generated. In classification.py it seems that octree already exists in the batch dictionary,

def forward(self, batch):
    octree, label = batch['octree'].cuda(), batch['label'].cuda()
    data = self.get_input_feature(octree)
    logits = self.model(data, octree, octree.depth)
    log_softmax = F.log_softmax(logits, dim=1)
    loss = F.nll_loss(log_softmax, label)
    pred = torch.argmax(logits, dim=1)
    accu = pred.eq(label).float().mean()
    return loss, accu

which looks like something from dataset.

But logically octree is generated upon the input point cloud, rather than being provided along with dataset. Could you explain the origin outcoming of octree, i.e. when/in which script was it generated and written into batch dict? Thanks a lot!

Hello, thank you very much for your interest in our work! We have referenced the design of our baseline, Octformer (https://arxiv.org/abs/2305.03045), for the specific implementation of our octree. The specific implementation code can be found in the ocnn library mentioned in requirement.txt. Indeed, I think as you said, the octree is added to the batch as a static dictionary.

My understanding is that the octree is statically defined in the batch you mentioned, and then instantiated with the input point cloud features in the following part in classification.py:

  def get_input_feature(self, octree):
    flags = self.FLAGS.MODEL
    octree_feature = ocnn.modules.InputFeature(flags.feature, flags.nempty)
    data = octree_feature(octree)
    return data

For further understanding, perhaps you could refer to Octformer.

[BEFLIPPED]

Hello, I have a confused question as follows.
In this project, How is the static dictionary batch defined and where is it defined ?
Thanks !