Point Cloud Denoising Competition Submission

This repository contains the cleaned source code and the released checkpoint for our Jittor point-cloud denoising competition submission.

Contents

• code/: training, inference, evaluation code, and the minimal configuration set used by the final model.
• checkpoints/seed2_ep87.pkl: released EMA checkpoint used for the final seed2 epoch-87 submission.
• report.pdf: add the competition report PDF exported from Overleaf before final publication.

Large official datasets, local caches, experiment logs, and generated submissions are intentionally excluded.

Environment

Install dependencies from code/requirements.txt in a Jittor-capable Python environment.

cd code
pip install -r requirements.txt

The original experiments used Jittor with CUDA. If your environment uses a custom Python executable, replace python below with that executable.

Data Layout

Place the official datasets under code/ when running locally:

code/dataset_train/
code/dataset_test_noisy/

The optional local validation split can be generated with:

cd code
python tools/build_testlike_validation.py \
  --train_root ./dataset_train \
  --output_root ./dataset_local_val_testlike \
  --datalist ./datalist/val_testlike.txt

Inference

Run the released checkpoint on the official test-noisy split:

cd code
python run.py --task configs/task/predict_seed2_ep87.yaml

Predictions are written to:

outputs/infer/test/seed2_ep87/

Build a submission zip from those predictions:

cd code
python tools/greedy_ensemble.py \
  --mode build \
  --test-outputs ../outputs/infer/test \
  --labels seed2_ep87 \
  --out-zip ../outputs/submissions/seed2_ep87.zip

Training

To retrain the released seed2 configuration:

cd code
python run.py --task configs/task/train_seed2.yaml

Checkpoints are written to:

outputs/iterpfn_seed2/checkpoints/

Notes

The repository intentionally does not include official datasets, local validation caches, experiment logs, or generated zip submissions.