Trouble using `.pkl` S3DIS data to train PVCNN + Semantic Segment ZED 2i .ply

[arnavk23]

I'm trying to train the PVCNN model using .pkl files (preprocessed from S3DIS data) and later perform semantic segmentation on .ply point clouds captured by a ZED 2i camera. However, the original prepare_data.py script fails, and the training pipeline expects .npy or .txt formats rather than .pkl.

What I Did:

Preprocessed .pkl Available:

I confirmed the following structure in my data:

with open('path/to/data.pkl', 'rb') as f:
    data = pickle.load(f)

# Sample output:
# coord: (N, 3)
# color: (N, 3)
# semantic_gt: (N, 1)
# instance_gt: (N, 1)
# normal: (N, 3)

Running prepare_data.py gave:

FileNotFoundError: .../Area_1/office_7/xyzrgb.npy

Because .npy files are expected, but my data is already processed and in .pkl.

Accidentally shadowed torch module:

I had a local file named torch.py, which led to:

ModuleNotFoundError: No module named 'torch.utils'; 'torch' is not a package

Fixed by renaming the file and clearing __pycache__.

Next:

I tried importing a s3dis_pkl_dataset class (not in repo), and got:

ModuleNotFoundError: No module named 's3dis_pkl_dataset'

Solution I Built:

Created a custom dataset class:

class S3DISPKLDataset(Dataset):
    ...
    def __getitem__(self, idx):
        with open(self.data_paths[idx], 'rb') as f:
            data = pickle.load(f)
        ...
        return {
            'coord': coord,
            'color': color,
            'semantic_gt': semantic_gt
        }

This allowed me to correctly load .pkl files and test DataLoader successfully.

Request for Maintainers:

1. Provide built-in support or documentation for using .pkl files as input.
2. Consider refactoring prepare_data.py to allow choosing between .txt, .npy, or .pkl.
3. Add a safeguard in README against naming files torch.py (common mistake for new users).
4. Confirm best practices for plugging in external .ply files (e.g. ZED 2i output) for inference.

Environment

• OS: Ubuntu 22.04
• Python: 3.13 (Miniconda)
• Torch: 2.7.1 (CUDA 12.6)
• PVCNN repo: latest from GitHub