CDSegNet Autoencoder Only

This mini-project extracts the conditional encoder–decoder branch from CDSegNet and turns it into a light-weight point-cloud semantic segmentation trainer for the Replica dataset. Diffusion components are deliberately omitted; the model focuses on hierarchical point feature encoding and decoding with skip connections.

Layout

• cdsegnet_autoencoder_only/ – reusable Python package exposing the Replica dataset loader, point autoencoder model, and utility helpers.
• train.py – entry point for launching training/evaluation runs.
• cache/ – autogenerated directory (when the training script runs) that hosts cached .npz copies of the Replica point clouds to avoid repeatedly parsing the large ASCII .pcd files.

Quickstart

cd /usr/project/CDSegNet_autoencoder_only
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python train.py \
  --data-root ../E_data/replica_dataset/preprocessed \
  --output-dir ./outputs \
  --train-scenes apartment_0,apartment_1 \
  --val-scenes apartment_2 \
  --epochs 50 \
  --batch-size 4

Key artefacts are written to outputs/:

• train_config.json – frozen copy of the CLI arguments.
• dataset_meta.pt – label mapping and class-name metadata.
• logs/metrics.txt – per-epoch training/validation metrics.
• checkpoints/latest.pt and checkpoints/best.pt – PyTorch checkpoints with model/optimizer state.

Training Notes

• The dataset class loads Replica point clouds once, caches them under cache/, and samples random 4,096-point subsets per __getitem__.
• Basic augmentations include random rotation around the gravity axis and Gaussian jitter. Validation runs without augmentation.
• The network implements a four-stage encoder/decoder with farthest-point sampling, local k-NN aggregation, and nearest-neighbour feature interpolation, mirroring the structure of CDSegNet's conditional branch.

Evaluating Custom Scenes

To run evaluation on a custom split, point the --val-scenes flag to the desired Replica scene names. The dataset automatically shares the label index mapping between training and validation splits to keep predictions aligned.