[PDF] 3d reconstruction from single image

What are the challenges of 3D reconstruction of a single image?

In general, the 3D reconstruction of a single image mainly has the following challenges: (1) shape complexity reconstruction of objects, (2) uncertainty reconstruction of objects, (3) reconstruction of fine-grained objects, (4) memory requirements and calculation time, (5) training datasets, (6) selectivity of 3D shape representations.

What is single view 3D reconstruction?

This problem is called single-view 3D reconstruction, and it is ill-posed since an image can be a projection of infinitely many 3D shapes. Interestingly, humans are very good at estimating the 3D structure of any known class object from a single image; we can even predict how it looks in unseen views.

Can a 3D image be reconstructed from an RGB image input?

Recently, significant advances have been made towards complete geometric reconstruction from an RGB image input [10, 33], as well as reconstruction of object instances detected in the image [15, 24, 28]; however, 3D semantic reconstruction and instance reconstruction have largely been considered independently.

Can 3D shape representation be used for 3D reconstruction?

There is no absolute standard for the use of 3D shape representation for the reconstruction of 3D objects from a single image. Different 3D representations have advantages and disadvantages. In addition, the training details of the 3D reconstruction network need to be further explored.