search

xxlong0 / Wonder3D

Single Image to 3D using Cross-Domain Diffusion for 3D Generation
https://www.xxlong.site/Wonder3D/
GNU Affero General Public License v3.0
4.78k stars 384 forks source link

some question about the camera pose #72

Open GFENGG opened 12 months ago

GFENGG commented 12 months ago

hello, when i use the 6 view config, i found the camera pose (elevations and azimuths) are:

elevations: [tensor(0.), tensor(-13.5356), tensor(-9.6611), tensor(29.9105), tensor(39.5716), tensor(21.3893)] azimuths: [tensor(0.), tensor(46.5458), tensor(97.0298), tensor(180.), tensor(277.0298), tensor(320.0548)]

(the above angle are changed from radian to degree by angle*180/np.pi) So, my question is why the elevations are not all zero, and why the azimuths are not [0, 45, 90, 180, -90, -45]

flamehaze1115 commented 12 months ago

Thanks for point it out! It is a bug. Since the model is trained using the fixed poses, the bug won't affact the final results, The wrong poses works like labels to distinguish different views.

GFENGG commented 12 months ago

Yes, the results can also work will on these views. And i have another question, can the model render color/normal maps under various camera poses, which are different from the given 6 camera poses? In other words, can the model support different camera pose embeddings as input?

flamehaze1115 commented 12 months ago

Yes, the results can also work will on these views. And i have another question, can the model render color/normal maps under various camera poses, which are different from the given 6 camera poses? In other words, can the model support different camera pose embeddings as input?

Yes. It is possible, but it requires training wonder3D using more views. We are working on this.