search

microsoft / Swin3D

A shift-window based transformer for 3D sparse tasks
MIT License
202 stars 19 forks source link

Problem when predicting sample data input.npz #18

Closed QuangHieu1704 closed 8 months ago

QuangHieu1704 commented 11 months ago

Hi, I am working with your source code and faced a problem. With your sample data "input.npz", I got below results. This result like a mess. I don't know where am I wrong, so can you please share with me your code for making prediction on own point cloud? Thank you very much.

Note: image1 is point cloud of "input.npz" image2 is prediction image3 is points of label "wall" image1 image2 image3

Yukichiii commented 10 months ago

Do you mean the predictions output by example.py? The model in example.py is not loaded with weights, but rather randomly initialized, which naturally give result like a mess.

QuangHieu1704 commented 10 months ago

Thanks for you reply. I have loaded the pretrained model. Actually, I didn't see the example.py script, I modified the examples/segmentation.py script. image

Yukichiii commented 10 months ago

If you load the ckpt provided in this repo, it may also be a mess. This is because the pretrained model is pretrained on the Structured3D Dataset, which is a synthetic dataset. But the input.npz is a sample from ScanNet which is a real dataset. The model need to be finetuned on ScanNet to get reasonable segmentation result. Our finetuned model is provided in this repo. If you want to forward and visualize the segmentation result, you can try this.