how to generate new dataset for TF records

eldar / differentiable-point-clouds

The reference implementation of "Unsupervised Learning of Shape and Pose with Differentiable Point Clouds"

MIT License

155 stars 11 forks source link

how to generate new dataset for TF records #1

Open netalkative opened 5 years ago

netalkative commented 5 years ago

I am trying "dpc" as a generator of point cloud for new category. To generate ground truth point clouds is obvious(dpc/densify/densify.py or chenhsuanlin/3D-point-cloud-generation's densify/densify.py).

But I cannot find a way of a new train/test/val dataset creation, especially sources of TF records. How do I prepare a train/test/val dataset to create TF records?

eldar commented 5 years ago

Hi @netalkative This is described in the instructions: https://github.com/eldar/differentiable-point-clouds/blob/master/README.md#prepare-training-data Please follow them carefully, there is a script that generated TF records out of the images.

netalkative commented 5 years ago

Thank you for replying, @eldar

I am afraid that "prepare-training-data" is not enough. "prepare-training-data" requires camera?.mat , depth?.png and render_?.png from -renders.tar.gz that you prepared in advance.

I want to know, how to create camera?.mat , depth?.png and render_?.png from my 3D model data. When you created -renders.tar.gz for airplane, car and chair, how did you? (i am worried about "dpc/data/splits/.file". which program creates it?)

Best Regards.

Yueeey commented 5 years ago

Hi @eldar

I'm afraid I have the similar questions. The numbers in the camera_.mat really confused me. Could you tell me the meanings of the extrinsic, K, pos, quat in the mat files? Thank you very much!