search

lijx10 / USIP

USIP: Unsupervised Stable Interest Point Detection from 3D Point Clouds. ICCV 2019
GNU General Public License v3.0
202 stars 28 forks source link

Kitti dataset, too big coordinates of keypoints in evaluation #17

Open anastasiia-kornilova opened 4 years ago

anastasiia-kornilova commented 4 years ago

Dear authors, this is a very cool work, thank you for sharing the code!

I tried to reproduce the results of work on the KITTI dataset. I used save_keypoints.py with path to the dataset from GDrive (numpy representations of PCs) and to the model best.pth from oxford/16384-512-k1k16-2d, The model works and generates keypoints, but in generated keypoints I obtained too huge values:

[[ 3.09118889e-02 -9.79395750e-01 -3.95004240e+07]
 [ 3.27124819e-02 -1.18620468e+00 -3.91861840e+07]
 [ 1.46894772e-02 -2.04260017e+00 -4.01456720e+07]
 [ 2.93604157e-04 -6.35924319e+03 -2.76408450e+06]
......

Could you, please, give me a tip where I can mistake?

lijx-nutonomy commented 4 years ago

One reason could be the incorrect reading of point clouds, e.g., reading as dtype=np.float32 or np.float64.

anastasiia-kornilova commented 4 years ago

Yes, that was the problem. Thank you a lot!