bad registration result on 3dmatch test data for demo.py

[dspeia]

HI, qinzheng

Thanks for your work.
according https://github.com/qinzheng93/GeoTransformer?tab=readme-ov-file#testing-on-your-own-data
I tested your data, the result is good.
CUDA_VISIBLE_DEVICES=0 python demo.py --src_file=../../data/demo/src.npy --ref_file=../../data/demo/ref.npy --gt_file=../../data/demo/gt.npy --weights=../../weights/geotransformer-3dmatch.pth.tar

But i got very bad visualization result (attachment pic) when use 3dmatch test data:
CUDA_VISIBLE_DEVICES=0 python demo.py --src_file=../../3dmatch/test/7-scenes-redkitchen_npy/cloud_bin_2.npy --ref_file=../../3dmatch/test/7-scenes-redkitchen_npy/cloud_bin_3.npy --gt_file=../../data/demo/gt.npy --weights=../../weights/geotransformer-3dmatch.pth.tar

the gt file and weight are same to before,
the src_file, ref_file is down_sample and save to npy from 3dmatch test-set(test/7-scenes-redkitchen):
voxel_down_pcd = pcd_data.voxel_down_sample(voxel_size=0.025)

I'm so confuse about this result, would you please tell me which ply file on 3dmatch is your data/demo/src.npy & dts.npy?
and are my preprocess correct (read ply->downsample->save npy)?

according #4 #12, i guess it will work just need to replace the input point cloud data of demo.py, but i can't get one time good result on test/7-scenes-redkitchen , do i need rewrite the dataset?

thanks for your any reply.

[dspeia]

After checking, the demo.py can output correct registration result, the reason is #107
Thanks for qin zheng's work again!

The reply for before questions, FYI:
Q: I'm so confuse about this result, would you please tell me which ply file on 3dmatch is your data/demo/src.npy & dts.npy?
A: the data/demo/src.npy & dts.npy is 3dmatch/test/7-scenes-redkitcheny/cloud_bin_0 and cloud_bin_6

Q: and are my preprocess correct (read ply->downsample->save npy)?
A: the steps is correct, we can get good registration base this preprocess.