Develop a deep learning model for learning the triangular face label assignment on AxisCube meshes, ensuring shape fidelity and structural validity requirements are met. -- 2
Budget: $750 – $1,500 USD
This project focuses on label learning tasks based on AxisCube structures. A AxisCube is a special geometric structure that aligns with main axes, where each face is perpendicular or parallel to the ±x/±y/±z axes, and adjacent faces are mutually perpendicular.
[图片]
The input consists of a triangular mesh and its existing label data (indicating the direction of each triangle face as ±x/y/z), and the output is a new label assignment. Existing metric functions are used to evaluate label quality, including:
- Fidelity: The degree of alignment between the label and the face normal.
- Compactness: The smoothness of region boundaries and the number of corner points.
- Validity Conditions: Whether the label combinations can legally form a AxisCube structure.
I've already written the code for the Soft Constraints (Fidelity and Compactness energy) and Hard Constraints (Validity conditions), and the dataset includes ground truth labels for each face.
We provide:
- 1000+ training data (expandable).
- Label visualization tool (polyscope).
- Validity check function.
- Code for label evaluation metrics.
You need to construct a deep learning model based on the existing data to output high-quality, valid label assignments.
2.2.3 Deep Learning for Label Learning (With Dataset)
- You need to learn good labels using AI methods (deep learning), as the labels directly determine the feasibility of constructing a AxisCube grid. The main task here is to train a model that predicts good labels.
- A dataset of over 1,000 examples is provided, sourced from a few different methods, with the best results chosen. Some manually improved results are also included.
The goal is to train a deep learning model to predict these labels, which can then be used for generating AxisCube grids automatically. Various deep learning frameworks for mesh generation can be referenced, and polyscope (a grid visualization library) is available for visualization.
[图片]
The input consists of a triangular mesh and its existing label data (indicating the direction of each triangle face as ±x/y/z), and the output is a new label assignment. Existing metric functions are used to evaluate label quality, including:
- Fidelity: The degree of alignment between the label and the face normal.
- Compactness: The smoothness of region boundaries and the number of corner points.
- Validity Conditions: Whether the label combinations can legally form a AxisCube structure.
I've already written the code for the Soft Constraints (Fidelity and Compactness energy) and Hard Constraints (Validity conditions), and the dataset includes ground truth labels for each face.
We provide:
- 1000+ training data (expandable).
- Label visualization tool (polyscope).
- Validity check function.
- Code for label evaluation metrics.
You need to construct a deep learning model based on the existing data to output high-quality, valid label assignments.
2.2.3 Deep Learning for Label Learning (With Dataset)
- You need to learn good labels using AI methods (deep learning), as the labels directly determine the feasibility of constructing a AxisCube grid. The main task here is to train a model that predicts good labels.
- A dataset of over 1,000 examples is provided, sourced from a few different methods, with the best results chosen. Some manually improved results are also included.
The goal is to train a deep learning model to predict these labels, which can then be used for generating AxisCube grids automatically. Various deep learning frameworks for mesh generation can be referenced, and polyscope (a grid visualization library) is available for visualization.