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.
e5e9d7b985eeee8b185125949e3e5ab8.mp4
- 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.
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! 🌟
For any inquiries or potential collaborations, please open an issue or contact:
[email protected]
This project is licensed under the Apache 2.0 License.