search

GAP-LAB-CUHK-SZ / LASA

CVPR2024 | LASA: Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset
85 stars 3 forks source link

Laser scan sampling #7

Open Vynikal opened 2 weeks ago

Vynikal commented 2 weeks ago

Hello, in the paper, on the page 4, you claim to have the laser clouds subsampled to 4mm. But in the downloaded LASA dataset, the resolution is only 4cm. That is unfortunately too sparse for my use-case. Is there any simple way to obtain the whole dataset with denser clouds? Thank you

Vynikal commented 2 weeks ago

And another question, the 'alignment.txt' rarely aligns anything, it is often just an eye matrix, or it has only small translations. When importing to cloudcompare, the instances are completely misaligned with the ARKitScenes, and rotated 90°. How to align the instances properly, am I doing something wrong?

hugoycj commented 3 days ago

@Vynikal Due to their large size (approximately 1GB per scene), sharing 4mm point clouds is challenging. We recommend downloading the raw laser data directly from the ArkitScenes repository and preprocessing it using Open3D.

For rotation problem, we are using opengl coordinate, while ARKitScenes use a different coordinate, we will share the visualization code in next few weeks for put LASA CAD into ARKitScenes

Vynikal commented 3 days ago

Thanks for the answer! Looking forward to the visualization code. But just a confirmation: in the current state of the dataset, there is no given transformation to align LASA models in the scene coordinates, right?