LOD-PCAC: Level-of-detail-based Deep Lossless Point Cloud Attribute Compression

• Python 3.9
• PyTorch 1.12.0
• Torchvision 0.13.0
• Torchaudio 0.12.0
• CUDA 11.6
• Pybind11
• TensorboardX
• Scikit-learn
• TQDM
• Einops
• PyTorch3D v0.7.2
• Ninja

You can download the dataset and pretrained models from the one of the following links:

https://pan.baidu.com/s/1J04bpSWZLiwYtzYDdXfNzw
Access Code: 84ac
https://1drv.ms/f/c/1345fa17d336f6c3/Eth9H98umstHl8T_pDmS2gYBFrsQnkIcy4FstDTKwE4Fug?e=dulVU8
Access Code: TCTW

After downloading, create three folders (train/, test/, dev/) in the project directory and extract the corresponding dataset archives into their respective folders.

We rely on the MPEG G-PCC codec to obtain LOD (Level of Detail) structures. For convenience, we provide a precompiled and modified codec for both Windows and Linux. Below are the steps for processing files in the train/ directory on a Linux system:

1. Place run-train.sh, tmc3_lod, and encoder.cfg into the train/ directory.
2. Execute run-train.sh to generate LOD information.
3. Modify the second line of pdata.py to "train_list.py".
4. Run pdata.py to validate and convert the LOD information into .npy format.

Repeat the same process for the test/ and dev/ directories.

For details on how MPEG generates LOD structures, refer to the buildPredictorsFast function in the MPEG source code:
buildPredictorsFast Function

To use the pretrained models, extract the pretrained model files and place them in the pretrain/ directory, and install the modified torchac. Then, you can run python encode.py and python decode.py for encoding and decoding.

To train the model from scratch, perform preprocessing as described above and then simply run python train.py