Xiaoyang Wu
Xiaoyang Wu
> line 701 point.serialization(order=self.order, shuffle_orders=self.shuffle_orders) This only computes each serialization's order but does not reorder the point cloud. > line 408-412: This part actually reorder the point cloud
Data Processing: https://github.com/Pointcept/Pointcept/blob/main/pointcept/datasets/preprocessing/scannet/dino/preprocess_dino_feature.py Segmentor: https://github.com/Pointcept/Pointcept/blob/main/pointcept/models/default.py#L97 I didn't write the README, yet the code is included here.
Hi, discussion here might be helpful (https://github.com/Pointcept/Pointcept/issues/284#issuecomment-2188177675). As the release model is trained with v1.5.1
Could you provide more detailed information to help me figure out the issue?
You can download my processed version for v1.5.2 from here (https://huggingface.co/datasets/Pointcept/s3dis-compressed)
Hi, train three PTv2 separately and vote on the prediction on one submission, which will deliver a better test performance. We encourage not using this trick for validation but only...
Hey, we don't have but can be easily fulfilled by creating one here (https://github.com/Pointcept/Pointcept/blob/main/pointcept/utils/scheduler.py#L35)
Typically, the currently default: Cosine with warmup is most popular, and model training is more stable now, so there is no need to carefully adjust the learning curve.
If you use PointGroup, it might need to adjust some parameters for your personal dataset.
That was strange. Could you provide your label mapping and parameters? (e.g. segment ignore index, instance ignore index)