CGAPoseNet+GCAN 📷

IEEE/CVF WACV 2024 - Waikoloa, HI

Geometry-Aware Camera Pose Regression with Geometric Clifford Algebra Networks

📌 Introduction

We introduce CGAPoseNet+GCAN, an enhanced version of CGAPoseNet for camera pose regression from RGB images. By incorporating a Geometric Clifford Algebra Network (GCAN), we create a geometry-aware pipeline that improves pose estimation accuracy while reducing computational overhead.

🔹 Key Features

• Unified Representation: Uses Clifford Geometric Algebra to model camera poses as motors, integrating quaternions and translation vectors into a single mathematical object.
• No Extra Scene Information Needed: Achieves competitive results without requiring 3D point clouds or expensive loss function tuning.
• Geometry-Aware Learning: GCAN refines motor proposals generated by an InceptionV3 backbone, ensuring mathematically meaningful predictions.
• State-of-the-Art Performance: Tested on 13 datasets, reducing:
  • Rotation error by 41% and translation error by 8.8% vs. CGAPoseNet.
  • Rotation error by 32.6% and translation error by 19.9% vs. the best PoseNet strategy.
• Efficient and Lightweight: Achieves these improvements with 4 million fewer trainable parameters compared to CGAPoseNet.

---

📁 Supplementary Materials

The supplementary materials contain two folders:

📂 code

• Includes the GitHub branch with the two Jupyter Notebooks and instructions for running them.

📂 results

• Contains:
  • Test MSEs
  • Positional and rotational errors for each of the 13 datasets.
• Note: Model weights are omitted due to size constraints but are available upon request.

Loss Terms:

• loss → CGAPoseNet+GCAN
• NB_loss → CGAPoseNet (without bottleneck)
• OG_loss → Original CGAPoseNet

---

📖 Citation

If you find this work useful, please cite:

@inproceedings{pepe2024cgaposenet+,
  title={Cgaposenet+ gcan: A geometric clifford algebra network for geometry-aware camera pose regression},
  author={Pepe, Alberto and Lasenby, Joan and Buchholz, Sven},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={6593--6603},
  year={2024}
}

---

🎯 CGAPoseNet+GCAN pushes the boundaries of camera pose regression with Geometric Algebra and Deep Learning. Feel free to reach out for any questions!🚀