szymanowiczs / splatter-image

Official implementation of `Splatter Image: Ultra-Fast Single-View 3D Reconstruction' CVPR 2024
https://szymanowiczs.github.io/splatter-image
BSD 3-Clause "New" or "Revised" License
795 stars 54 forks source link

How-To use the splatter-image on own images #21

Closed DiamondGlassDrill closed 4 months ago

DiamondGlassDrill commented 6 months ago

Just from the reading perspective an awesome work! Congrats.

Yet I couldn't step into training on cars training dataset, but rather to do the cars only, I would love todo my own images.

Could u provide the steps in a short How-To to be taken that I can use any image to 3D?

Thanks in advance for the support.

johnren-code commented 6 months ago

same question, how to train with my own images with rgbs and camera poses?

Ira569 commented 5 months ago

my advice, maybe you could refer to srn.py,and add your code to make your dataset adopt the srn_car dataset format. I think the main challenge is to find out the meaning of transform matrix and projection matrix in srn.py Notes that position vectors are row vectors,and Gaussian rasterizer assumes row-major order of rigid body transform matrices

vector: (x,y,z,1) transform matrixs: |R 0| |T 1|

szymanowiczs commented 4 months ago

Indeed - you need to modify the dataloader. Thanks @Ira569 for pointing this out - for more details about camera conventions see the Readme.