3d mesh reconstruction of a bike is very bad compared to an airplane

[monacv]

Thanks a lot for open-sourcing your source code.

When I tried your demo code on an image of a bike, the results were not good at all compared to when tried on image of an airplane.

Do you have any suggestions for improvement?

However for airplane I have these

[MrMois]

There are two things you could try:

1. Train on a dataset specific for bike models. ShapeNet contains significanty more plane models (4045) than bike models (59): https://shapenet.org/taxonomy-viewer
2. Crop the image, so that just the bike is in focus. Compared to the plane image, there is quite alot of other stuff in the image like people and houses. There might also be a network available for this.

[monacv]

@MrMois thanks a lot for your feedback.

So I downloaded this bike image which is transparent and only contains one bike and still the image is terribly rendered.

Do you know which dataset would have 3D bikes that would be similar in features to ShapeNet or could you guide me how to create my own such dataset?

[Msj5]

Hi, This model just not very rubust to complex images. You can try it with some sample and smooth image like chair .
crop the image and convert it to .png(alpha = 0) @monacv

[topinfrassi01]

Since the deformation starts from an ellipsoid, it cannot generalize to meshs that aren't from the same genus, meaning that meshs with holes in it like the bike won't be reconstructed properly.

[harisraharjo]

To be able to train this model with specific dataset, do I need to prepare an image with multiple views or just one view?