NOVA: Next-step Open-Vocabulary Autoregression for 3D Multi-Object Tracking in Autonomous Driving 🌟

March 9, 2026 · View on GitHub

License Status Python PyTorch arXiv

NOVA is a novel 3D tracking framework that redefines object perception through the Autoregressive Paradigm. By formulating tracking as a sequential prediction task within an Open-Vocabulary state space, NOVA achieves unprecedented generalization across unseen categories in complex 3D environments.

🖼️ Teaser

NOVA Teaser

🎥 Demo

📅 Roadmap

  • Technical Report/Paper: Detailed methodology and experimental results. (arXiv)
  • Inference Code: Core implementation of the NOVA architecture.
  • Pre-trained Weights: Model checkpoints trained on large-scale 3D datasets.
  • Evaluation Suite: Scripts for benchmarking on standard 3D tracking datasets.

🚀 Coming Soon

We are currently cleaning up the codebase and preparing comprehensive documentation. The full source code, pre-trained models, and usage examples will be released shortly.

Stay tuned by Starring the repository! 🌟

✉️ Contact

For any inquiries or potential collaborations, please open an issue or contact: xifen527@163.com

📄 License

This project is licensed under the Apache 2.0 License.