facebookresearch / vggsfm

VGGSfM: Visual Geometry Grounded Deep Structure From Motion
Other
830 stars 52 forks source link

Poor reconstruction on textureless surfaces #49

Closed LaFeuilleMorte closed 6 hours ago

LaFeuilleMorte commented 1 month ago

Hi, thanks for your brilliant work! I was using your HF demo to reconstruct a 1 minute long videos from an indoor scene. But the reconstruction results and the pose were very poor. Especially on the textureless surfaces (floor, wall, ceiling, doors, windows). There're rarely any points on the textureless surfaces. Is there any parameters I can tune to mitigate this problem. My parameter was using the default:

image

Thanks in advance for your reply

jytime commented 1 month ago

Hey @LaFeuilleMorte ,

Our hugging face demo is only a very basic illustration. A lot of features have not been supported in our hugging face demo.

For your case, after you install vggsfm as guided, please try:

python demo.py SCENE_DIR =/YOUR/PATH shared_camera=True camera_type=SIMPLE_RADIAL query_method=aliked
LaFeuilleMorte commented 1 month ago

Hey @LaFeuilleMorte ,

Our hugging face demo is only a very basic illustration. A lot of features have not been supported in our hugging face demo.

For your case, after you install vggsfm as guided, please try:

python demo.py SCENE_DIR =/YOUR/PATH shared_camera=True camera_type=SIMPLE_RADIAL query_method=aliked

Thanks for your reply! What's the function of camera type in this method? Would it be ok if I set camera_type=PINHOLE ?

jytime commented 1 month ago

Hi now we support

SIMPLE_ PINHOLE and SIMPLE_RADIAL

For SIMPLE_ PINHOLE, the camera is formulated by (f, cx, cy). For SIMPLE_RADIAL, it is (f, cx, cy, k)