CWPFormer is a novel Visual Place Recognition (VPR) framework based on the Cross-weight Pyramid Transformer, designed for efficient robot navigation and localization tasks. This framework leverages advanced attention mechanisms like Cascaded Hash Attention (CHA) and a Bayesian learning scheme to enhance recognition accuracy and reduce computational redundancy, making it suitable for real-world robotics applications.

Features

• Cross-weight Pyramid Transformer (CWPFormer): Extracts multi-scale visual features with efficient attention-based mechanisms.
• Cascaded Hash Attention (CHA): Reduces computational cost while improving attention diversity.
• Bayesian Learning: Enhances the place matching accuracy by learning compact hash codes.
• Real-time VPR Performance: Designed for autonomous vehicles, drones, and other perception-based unmanned systems.

Results

Requirements

The project requires Python 3.7+ and PyTorch 1.10+. All necessary dependencies are listed in the requirements.txt file.

Installation

Clone the repository:

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

Install the required Python packages:

pip install -r requirements.txt

Datasets Make sure you download the data set for training and evaluation. We used the following data sets:

• KITTI: Visual data set for robot positioning and navigation.
• Nordland: Challenge datasets for seasonal and environmental changes.
• EuRoC: Micro Air Vehicle dataset containing visual SLAM related data.

After downloading the dataset, extract it and place it in the specified directory.

Training To train the CWPFormer model on a dataset, run:

python train.py --data_dir /path/to/dataset --epochs 10 --batch_size 64 --learning_rate 1e-4

This will train the model using the settings defined in config.py. The model checkpoints will be saved in the checkpoints/ directory, and logs will be available in the logs/ directory.

Testing After training, you can test the model's performance on a test dataset by running:

python test.py --model_checkpoint /path/to/checkpoint --test_dir /path/to/test_data

This will load the latest model checkpoint from checkpoints/ and evaluate it on the test data, reporting the accuracy.