LiDAR Place Recognition in Vineyards with Matryoshka Learning
A new lightweight deep-learning method, MinkUNeXt-VINE, outperforms state-of-the-art approaches for place recognition in vineyard environments. The method uses low-cost sparse LiDAR inputs and lower-dimensionality outputs for high efficiency in real-time scenarios. It leverages Matryoshka Representation Learning multi-loss and pre-processing. Results are validated on two extensive long-term vineyard datasets with different LiDAR sensors, demonstrating efficacy. The study addresses localization challenges in unstructured agricultural settings lacking distinctive landmarks.
Key facts
- MinkUNeXt-VINE is a lightweight deep-learning method for place recognition in vineyards.
- It surpasses state-of-the-art methods in vineyard environments.
- Uses low-cost, sparse LiDAR inputs and lower-dimensionality outputs.
- Employs Matryoshka Representation Learning multi-loss approach.
- Validated on two long-term vineyard datasets with different LiDAR sensors.
- Addresses localization challenges in unstructured agricultural environments.
- Prioritizes enhanced performance and high efficiency in real-time scenarios.
- Includes a comprehensive ablation study on various evaluation cases.
Entities
—