Hi TensorFlow Graphics Team / Authors,
I am trying to run the "From Points to Multi-Object 3D Reconstruction" (CVPR 2021) project located in the tensorflow_graphics/projects/points_to_3Dobjects directory of this repository.
I have followed the preprocessing steps (downloading CoReNet and ShapeNet, running preprocess_shapenet.py and preprocess_clusters.py) as described in the README found in the original francisengelmann/points2objects repository (which seems to contain only preprocessing code).
I then looked at the points_to_3Dobjects project within this (tensorflow/graphics) repository, assuming it contains the full implementation. The README here confirms this is the project related to the paper, but also notes it is "Work in progress".
The training script (train_multi_objects/train.py) clearly requires data in TFRecord format via the --tfrecords_dir flag. I found the data_preparation/data.proto file, which seems to define the TFRecord structure, and data_preparation/extract_protos.py, which contains functions (decode_bytes_multiple) used by train.py to parse these TFRecords.
However, I cannot find the script or instructions detailing how to generate these TFRecords from the original CoReNet dataset (.npz files) and the associated ShapeNet/SDF/point cloud data. I have checked the data_preparation/ and tools/ directories within the project, as well as the READMEs and the original paper, but this step seems to be missing.
I noticed Issue #467 ("NASA: generating input tfrecord files") from December 2020 asked a similar question regarding TFRecord generation for this project (referencing the paper's sampling methods) but appears unanswered.
Could you please provide guidance on how to generate the TFRecords required by train_multi_objects/train.py?
Is the generation script located elsewhere in the repository?
Was it perhaps omitted from the public release?
Is there an alternative data loading pipeline recommended for this implementation?
Any help or pointers would be greatly appreciated! Thank you.
Hi TensorFlow Graphics Team / Authors,
I am trying to run the "From Points to Multi-Object 3D Reconstruction" (CVPR 2021) project located in the tensorflow_graphics/projects/points_to_3Dobjects directory of this repository.
I have followed the preprocessing steps (downloading CoReNet and ShapeNet, running preprocess_shapenet.py and preprocess_clusters.py) as described in the README found in the original francisengelmann/points2objects repository (which seems to contain only preprocessing code).
I then looked at the points_to_3Dobjects project within this (tensorflow/graphics) repository, assuming it contains the full implementation. The README here confirms this is the project related to the paper, but also notes it is "Work in progress".
The training script (train_multi_objects/train.py) clearly requires data in TFRecord format via the --tfrecords_dir flag. I found the data_preparation/data.proto file, which seems to define the TFRecord structure, and data_preparation/extract_protos.py, which contains functions (decode_bytes_multiple) used by train.py to parse these TFRecords.
However, I cannot find the script or instructions detailing how to generate these TFRecords from the original CoReNet dataset (.npz files) and the associated ShapeNet/SDF/point cloud data. I have checked the data_preparation/ and tools/ directories within the project, as well as the READMEs and the original paper, but this step seems to be missing.
I noticed Issue #467 ("NASA: generating input tfrecord files") from December 2020 asked a similar question regarding TFRecord generation for this project (referencing the paper's sampling methods) but appears unanswered.
Could you please provide guidance on how to generate the TFRecords required by train_multi_objects/train.py?
Is the generation script located elsewhere in the repository?
Was it perhaps omitted from the public release?
Is there an alternative data loading pipeline recommended for this implementation?
Any help or pointers would be greatly appreciated! Thank you.