[GSOC] Pix3D dataset implementation

tensorflow / datasets

TFDS is a collection of datasets ready to use with TensorFlow, Jax, ...

https://www.tensorflow.org/datasets

Apache License 2.0

4.29k stars 1.54k forks source link

[GSOC] Pix3D dataset implementation #1756

Open aradhyamathur opened 4 years ago

aradhyamathur commented 4 years ago

I was successfully able to partially create a new dataset for Pix3D, link to commit Pix3D contains images, masks and models. I was able to create the split generator for images and masks since they both can be treated as Image features. However, currently there are no implementations for 3D Models as features. I would like to get some guidance on how to proceed about it, perhaps could I start by creating custom features for mesh comprising of edges and vertices ?

Eshan-Agarwal commented 4 years ago

link to commit is not correct I think it is for wiki40b.py can you please check

aradhyamathur commented 4 years ago

link to commit is not correct I think it is for wiki40b.py can you please check

thanks, updated

Conchylicultor commented 4 years ago

Thanks for reporting. Someone is currently adding a tfds.features.TriangleMesh. I believe it will be merged next week.

It would be equivalent to:

tfds.features.TriangleMesh == FeatureDict({
      'vertices': feature_lib.Tensor(shape=(None, 3), dtype=tf.float64),
      'faces': feature_lib.Tensor(shape=(None, 3), dtype=tf.uint64),
})

As input, it accept for encoding:

str: path to a {obj,stl,ply,glb} triangle mesh.
trimesh.base.Trimesh: A triangle mesh object.
trimesh.base.Scene: A scene object containing multiple TriangleMesh objects.
- A dict {'vertices': [...], 'faces': [...]}