search

idsia-robotics / RandLA-Net-pytorch

Our PyTorch implementation of RandLA-Net (https://github.com/QingyongHu/RandLA-Net)
26 stars 5 forks source link

标签未对齐 #1

Open obouluh opened 3 years ago

obouluh commented 3 years ago

并非每个id的点云都能包含所有的类型的标签，训练时配置项默认设定的label数量为5,一旦出现不一致的情况，就会警示⚠️ f"self.cfg['num_classes'] {self.cfg['num_classes']} is different" \ AssertionError: self.cfg['num_classes'] 5 is differentfrom len(pc_labels) 3

Gabry993 commented 3 years ago

Hi. Yes, this is not well documented, but the idea is that the first point cloud provided in the point cloud list must contain all the possible labels (i.e. len(pc_labels) == self.cfg['num_classes']). Then, other point clouds must contain a subset of those. This is not ideal, but it was convenient for our own setup. Changing this behavior shouldn't be too complicated. I'll try to change it if I find the time to do it :)