Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Low Cost, High Efficiency: LiDAR Place Recognition in Vineyards with Matryoshka Representation Learning

Judith Vilella-Cantos, Mauro Martini, Marcello Chiaberge, Mónica Ballesta, David Valiente

TL;DR

The paper tackles LiDAR-based place recognition for autonomous navigation in vineyards, where repetitive plant structures and seasonal changes challenge localization. It introduces MinkUNeXt-VINE, a lightweight LPR pipeline that prunes the backbone and uses Matryoshka Representation Learning to produce a compact 192-dimensional descriptor from sparse LiDAR data, with intensity processing and GeM pooling to boost robustness. Through extensive ablations and cross-season validations on the Bacchus Long-Term and TEMPO-VINE datasets, the method demonstrates strong performance gains over urban baselines, even with low-cost sensors like Livox, and maintains real-time inference suitable for onboard deployment. These results underscore the value of architecture pruning, careful input preprocessing, and multi-loss training for robust, efficient vineyard localization, with potential extensions to multi-modal sensing and field-deployable systems.

Abstract

Localization in agricultural environments is challenging due to their unstructured nature and lack of distinctive landmarks. Although agricultural settings have been studied in the context of object classification and segmentation, the place recognition task for mobile robots is not trivial in the current state of the art. In this study, we propose MinkUNeXt-VINE, a lightweight, deep-learning-based method that surpasses state-of-the-art methods in vineyard environments thanks to its pre-processing and Matryoshka Representation Learning multi-loss approach. Our method prioritizes enhanced performance with low-cost, sparse LiDAR inputs and lower-dimensionality outputs to ensure high efficiency in real-time scenarios. Additionally, we present a comprehensive ablation study of the results on various evaluation cases and two extensive long-term vineyard datasets employing different LiDAR sensors. The results demonstrate the efficiency of the trade-off output produced by this approach, as well as its robust performance on low-cost and low-resolution input data. The code is publicly available for reproduction.

Low Cost, High Efficiency: LiDAR Place Recognition in Vineyards with Matryoshka Representation Learning

TL;DR

The paper tackles LiDAR-based place recognition for autonomous navigation in vineyards, where repetitive plant structures and seasonal changes challenge localization. It introduces MinkUNeXt-VINE, a lightweight LPR pipeline that prunes the backbone and uses Matryoshka Representation Learning to produce a compact 192-dimensional descriptor from sparse LiDAR data, with intensity processing and GeM pooling to boost robustness. Through extensive ablations and cross-season validations on the Bacchus Long-Term and TEMPO-VINE datasets, the method demonstrates strong performance gains over urban baselines, even with low-cost sensors like Livox, and maintains real-time inference suitable for onboard deployment. These results underscore the value of architecture pruning, careful input preprocessing, and multi-loss training for robust, efficient vineyard localization, with potential extensions to multi-modal sensing and field-deployable systems.

Abstract

Localization in agricultural environments is challenging due to their unstructured nature and lack of distinctive landmarks. Although agricultural settings have been studied in the context of object classification and segmentation, the place recognition task for mobile robots is not trivial in the current state of the art. In this study, we propose MinkUNeXt-VINE, a lightweight, deep-learning-based method that surpasses state-of-the-art methods in vineyard environments thanks to its pre-processing and Matryoshka Representation Learning multi-loss approach. Our method prioritizes enhanced performance with low-cost, sparse LiDAR inputs and lower-dimensionality outputs to ensure high efficiency in real-time scenarios. Additionally, we present a comprehensive ablation study of the results on various evaluation cases and two extensive long-term vineyard datasets employing different LiDAR sensors. The results demonstrate the efficiency of the trade-off output produced by this approach, as well as its robust performance on low-cost and low-resolution input data. The code is publicly available for reproduction.
Paper Structure (20 sections, 5 equations, 8 figures, 10 tables)

This paper contains 20 sections, 5 equations, 8 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Place recognition in agricultural environments
  4. LiDAR place recognition
  5. Methodology
  6. Input preprocessing
  7. MinkUNeXt-VINE
  8. Learning with a multi-loss approach
  9. Datasets
  10. Bacchus Long-Term Dataset
  11. TEMPO-VINE dataset
  12. Experimental setup
  13. Ablation study
  14. Dimensionality study
  15. Impact of design choices
  16. ...and 5 more sections

Figures (8)

  • Figure 1: MinkUNeXt-VINE's proposed architecture. The parent backbone has been pruned to improve efficiency and performance in unstructured settings.
  • Figure 2: Extension of each of the vineyard considered within the BLT datasets polvara2024bacchus. a) Ktima Gerovassiliou (Greece). b) Riseholme (United Kingdom).
  • Figure 3: Comparison between the two LiDAR sensors used in the TEMPO-VINE dataset collection. Left: Point cloud captured with the Livox. Right: Point cloud captured with the Velodyne. The vertical FoV is annotated next to each sensor, while the horizontal FoV is 360$^{\circ}$ for both.
  • Figure 4: Extension of the field where the TEMPO-VINE dataset martini2025tempo was recorded.
  • Figure 5: Recall@N graph for the different design decisions tested.
  • ...and 3 more figures