3DTopia / LGM

[ECCV 2024 Oral] LGM: Large Multi-View Gaussian Model for High-Resolution 3D Content Creation.
https://me.kiui.moe/lgm/
MIT License
1.62k stars 105 forks source link

Training view angle details #22

Open tuna0711 opened 7 months ago

tuna0711 commented 7 months ago

Hello, it is a really nice work! May I ask about the exact elevation and azimuths angles for your 100 camera views used in training? Thanks!

ashawkey commented 7 months ago

Hi, it's basically sampled from a spiral path on a sphere surface: image

tuna0711 commented 7 months ago

Thanks! Looks like the camera elevation simply goes from 90 to -90, while the azimuth goes from 0 to 360 around 5 rounds? Also, during training, are there any issues when the randomly selected input views are too different from other views that used for supervision?