KTPFormer: Keypoint Transformer for Camera Matrix Prediction

A deep learning model that predicts camera matrices from 2D keypoints using a transformer-based architecture with graph convolutions.

Features

• Graph-based keypoint processing
• Transformer architecture for sequence modeling
• Combined Frobenius norm and reconstruction loss
• TensorBoard visualization support
• MongoDB integration for data management

Project Structure

• model/: Contains the model architecture and training logic.
• data/: Handles data loading and preprocessing.
• utils/: Utility functions for data visualization and logging.
• runs/: Stores model checkpoints and TensorBoard logs.
• weights/: Saved model weights.
• Readme.md: This file.

Getting Started

1. Clone the repository:

git clone https://github.com/your-username/ktpformer.git
cd ktpformer

2. Install dependencies:

pip install -r requirements.txt

3. Set up your environment variables:

cp .env.example .env

4. Run the training script:

python train.py