search

graphdeco-inria / gaussian-splatting

Original reference implementation of "3D Gaussian Splatting for Real-Time Radiance Field Rendering"
https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/
Other
14.05k stars 1.82k forks source link

How can i improve my result in driving scene #987

Open zzz5y opened 1 month ago

zzz5y commented 1 month ago

Hi! Thank you for your amazing work! Hi communit,i am dealing driving scenes reconstrcuctions.In 3dgs rebuild i got a bad result. Are there any idea to improve my work? Very appreciate for you kindly help! I train kitti360 in 60k or 100k iterations. and I use colmap to generate the points3d.ply instead of the dataset ply here are my bad results Hopefully to receive your ideas!Thank you

img_v3_02em_b3d7d8cc-db19-4570-a04e-25dd9a3f182g img_v3_02em_ebd95077-5e68-4635-b9ba-92c093d6421g img_v3_02em_c698b11e-fb4b-46fd-bbae-7f5d117c05cg img_v3_02em_b0115343-90b7-48f8-bb29-363d0ff2309g

jaco001 commented 1 month ago

In that such a reconstruction - you got only bearable result if you move reconstruction camera same way as you drive a car. And this isn't algorytm fail, but lack of useful data.

How to improve:

This way you got more coverage of the scene and your end effect will be more coherent with better perspective (less floaters, good flat surfaces etc.)

zzz5y commented 1 month ago

In that such a reconstruction - you got only bearable result if you move reconstruction camera same way as you drive a car. And this isn't algorytm fail, but lack of useful data.

How to improve:

  • add side cameras and longer video.
  • drive back same way to take 'counter' video

This way you got more coverage of the scene and your end effect will be more coherent with better perspective (less floaters, good flat surfaces etc.)

Thank you for your advice!It helps me a lot!

leblond14u commented 3 weeks ago

Hi,

I'm facing a similar issue with relatively low-res driving images where I get a sort of trail of blurry gaussians that densify directly in front of my render camera. All the initial gaussians generated from my initial point-cloud that are in the FoV of my images are pruned in the process (not the ones outside of my camera FoV weirdly enough).

Does anyone already encountered this kind of issue ?

Thanks in avance, Best,

AsherJingkongChen commented 4 days ago

Try lower position learning rate (fine-grained textures) and lower densification interval (more points) in argument classes.

However, you need more camera/view data.