CNPM

This repository contains implementation of Color Neural Parametric Models. A learnable parametric model for object shape and color representations.

Repository structure

This repo contains two implementation for CNPM model, their source roots can me found in following directories:

• src/cnpm
• src/cnpm_v2

Each implementation root contains 4 modules:

• src/{model_name}/callbacks Are responsible for PyTorchLightning callbacks used by training loop
• src/{model_name}/data Implementation of DataSets and DataModules
• src/{model_name}/model Actual implementation of a model
• src/{model_name}/traning Training methods that setup callbacks, loggers and run training loop

Training

Training configuration for each of the model can be found in corresponding root repositories. In order to run training, you should run following command:

python3 train.py --model cnpm --seed 42

To get full list of commands:

python3 train.py --help

Dataset

Datasets are publicly available by the following links:

• CNPM-V2: https://disk.yandex.com/d/mqaClruNZQwhsA
• CNPM: https://disk.yandex.com/d/mqaClruNZQwhsA

Optimization

Once the model is trained, you can use optimization scripts from src/cnpm/scripts/optimization, they generate artifacts with optimization process, that later can be used for statistical analysis.

An example of optimization process made on cloud points from multiple renders: