Deep Learning-Based 3D Point Cloud Denoising

Method Design and Implementation — Track A

This repository contains the Track A implementation submitted to the Sixth CG Graphics AI Challenge. It uses three released Jittor checkpoints and a two-stage weighted ensemble for whole-cloud denoising.

The submitted Track A result was Total 81.05, CD 69.86, and P2S 92.25. Reproducing the result requires the official data, the released checkpoints, and the settings below.

Contents

• Requirements
• Installation
• Data Preparation
• Training
• Inference
• Results and Reproducibility
• License and Acknowledgement

Requirements

Component	Tested version
OS	Ubuntu 22.04
GPU	NVIDIA RTX 4090
CUDA	12.4
Python	3.10
Jittor	1.3.10.0
Compiler	g++ 10

Released inference checkpoints:

checkpoints/control_e32_step_00050688.pkl
checkpoints/control_e34_step_00053856.pkl
checkpoints/noisyfit_e34_step_00053856.pkl

Verify them before use:

(cd checkpoints && sha256sum -c SHA256SUMS)

Installation

Create the environment from the repository root. Jittor compiles and caches CUDA operators on first use, so JITTOR_HOME must be writable. Change CUDA_HOME if CUDA is installed elsewhere.

conda env create -f environment.yaml
conda activate jittor

export cc_path="$(command -v g++)"
export nvcc_path=/usr/local/cuda-12.4/bin/nvcc
export CUDA_HOME=/usr/local/cuda-12.4
export JITTOR_HOME=/absolute/writable/path/jittor_cache
export conv_opt=1
mkdir -p "$JITTOR_HOME"

python scripts/check_environment.py --environment-file environment.yaml --cc-path "$cc_path" --nvcc-path "$nvcc_path" --cuda-home "$CUDA_HOME"

export WORK_ROOT=/absolute/writable/path/point_denoising_work
mkdir -p "$WORK_ROOT"

Continue only after the checker prints ENVIRONMENT_OK. Re-export the five environment variables in each new terminal before training or inference.

Data Preparation

The official data are not redistributed here. Set the archive paths supplied by the competition organizer, then prepare the training and noisy-test trees.

export TRAIN_DATASET=/path/to/dataset_train.tar.gz
export TEST_DATASET=/path/to/dataset_test_noisy.zip

python scripts/prepare_data.py --train-archive "$TRAIN_DATASET" --test-archive "$TEST_DATASET" --output-dir data/prepared --split-dir data/splits/seed_20260726_v2 --seed 20260726 --val-ratio 0.05

python scripts/build_all_train_split.py --source-split data/splits/seed_20260726_v2 --output-dir data/splits/all_15833_seed20260726_v1

python scripts/build_train_surface_cache.py --mesh-root data/prepared/train/dataset_train/shapenet --train-split data/splits/all_15833_seed20260726_v1/train.json --output-dir data/train_cache/surface50000_seed20260726_n15833_v1 --select-count 15833 --num-points 50000 --seed 20260726 --workers 16

The deterministic clean-surface cache contains 15,833 arrays of shape (50000, 3). It is required for training, but not for checkpoint inference.

Training

The two training arms use all 15,833 meshes, a patch size of 1000, batch size 10, 34 epochs, a learning rate of 5e-4, and seed 20260726. They can run on separate GPUs.

CUDA_VISIBLE_DEVICES=0 python -u scripts/train.py --config configs/train/pgd_all15833_starter_laplace_condgate_huber75_d010_e34_seed20260726.yaml --run-dir "$WORK_ROOT/runs/control"

CUDA_VISIBLE_DEVICES=1 python -u scripts/train.py --config configs/train/pgd_all15833_gate_e34_noisyfit_cube24_seed20260726.yaml --run-dir "$WORK_ROOT/runs/noisyfit"

Each epoch has 1,584 steps. Resume an interrupted run by appending –resume /path/to/checkpoint.pkl to its original command.

Inference

The commands below reproduce the released three-checkpoint inference recipe. The expected config digests allow checkpoint replay from data/prepared/test without building the multi-gigabyte training cache.

python scripts/denoise.py --config configs/train/pgd_all15833_starter_laplace_condgate_huber75_d010_e34_seed20260726.yaml --checkpoint checkpoints/control_e32_step_00050688.pkl --expected-config-sha256 981f6edbafc1bdc312c6ae3dfa7538965315b70cda2f85bd1bd6fbad27428c27 --input-dir data/prepared/test --output-dir "$WORK_ROOT/pred/control_e32" --patch-size 1000 --seed-k 6.0 --patch-batch-size 20 --niters 1 --normalization-mode noisy_max --fusion-mode hard_best --iteration-damping 1.0

python scripts/denoise.py --config configs/train/pgd_all15833_starter_laplace_condgate_huber75_d010_e34_seed20260726.yaml --checkpoint checkpoints/control_e34_step_00053856.pkl --expected-config-sha256 981f6edbafc1bdc312c6ae3dfa7538965315b70cda2f85bd1bd6fbad27428c27 --input-dir data/prepared/test --output-dir "$WORK_ROOT/pred/control_e34" --patch-size 1000 --seed-k 6.0 --patch-batch-size 20 --niters 1 --normalization-mode noisy_max --fusion-mode hard_best --iteration-damping 1.0

python scripts/denoise.py --config configs/train/pgd_all15833_gate_e34_noisyfit_cube24_seed20260726.yaml --checkpoint checkpoints/noisyfit_e34_step_00053856.pkl --expected-config-sha256 9e6cf10fc272a92522fd979369137ca8b7778428906c90e2b0e7fea213e7992e --input-dir data/prepared/test --output-dir "$WORK_ROOT/pred/noisyfit_e34" --patch-size 1000 --seed-k 6.0 --patch-batch-size 20 --niters 1 --normalization-mode noisy_max --fusion-mode hard_best --iteration-damping 1.0

Fuse the three prediction directories in the submitted order.

python scripts/ensemble_predictions.py --pred-a "$WORK_ROOT/pred/control_e32" --pred-b "$WORK_ROOT/pred/control_e34" --output-dir "$WORK_ROOT/pred/control_fused" --beta 0.25 --recipe-output "$WORK_ROOT/pred/control_recipe.json"

python scripts/ensemble_predictions.py --pred-a "$WORK_ROOT/pred/control_fused" --pred-b "$WORK_ROOT/pred/noisyfit_e34" --output-dir "$WORK_ROOT/pred/final" --beta 0.875 --recipe-output "$WORK_ROOT/pred/final_recipe.json"

The final denoised arrays and the ensemble recipe are written to $WOR{K}_{R}OOT\mathrm{/}pred\mathrm{/}finalandWORK\_ROOT/pred/final\; and$WORK_ROOT/pred/final_recipe.json. The RTX 4090 setting uses patch_batch_size=20; reduce it only if GPU memory is insufficient.

Results and Reproducibility

Item	Submitted setting
Control checkpoints	e32 and e34
Noisyfit checkpoint	e34
Patch size	1000
Patch coverage	seed_k=6, hard_best
First ensemble	control-e32 + 0.25 control-e34
Final ensemble	control-fused + 0.875 noisyfit-e34
Training seed	20260726

CD is bidirectional Chamfer distance and P2S is point-to-surface distance. A local rerun can vary numerically across CUDA and Jittor environments; it is not an independent verification of the online leaderboard score.

License and Acknowledgement

See LICENSE and NOTICE. The method is informed by Guiding Point Cloud Denoising with Learned Structural Priors.