Dense Depth: 3D from a single image

3D scan is a very expensive technology but you can reach good result with MeshRoom but you need to much photos. But in some simple cases you can reach good result also with only one image and the deapth map. The camera with depth are more accessible, but you can have the same results without it and the DenseDepth.

I do my work at this git project that is a for of the offical one with modifications: trained model with pytorch instead of tensorflow, save model to .obj format with demo_torch.py but also a notebook version.

Notebook Zero config

You can try on your favorite image by using this colab notebook: you need to edit 5 block to download your image and 6 block with the name of your image,

after you can run all boxes and on the left panel you can download the model.obj

You can also ignore the 5 step and charge your image by drag and drop it in the Files view where you can find also the model.obj

Results

Input image

All results 3D reconstructed images seams perfect from a point of view.

But the results is not perfect.

But from a single image this is a very good work, with some 3D editing you can reach good results from a single photo.

Conclusions

This is a very good results but it can be integrated in meshroom and I’m trying to do it here , if someone can get help to port this AI into meshroom.

Enjoy with your 3D modelling and contribute to open source, because it is Art!

Linux user and Open Source fun. Deep learning PhD Student, Full stack web developer, Mobile developer