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

AI summary

Overview

• Research area: Robotics / computer vision — LiDAR place recognition (LPR) for mobile-robot localization in agricultural environments, specifically vineyards.
• Technical level: Advanced. The paper assumes familiarity with 3D sparse convolutions, metric-learning loss functions (average-precision losses, contrastive/triplet losses), retrieval metrics (Recall@1, Recall@1%), and descriptor embedding design.
• Scope in one sentence: The paper proposes MinkUNeXt-VINE, a pruned and lightweight LiDAR place recognition network trained with a Matryoshka Representation Learning multi-loss scheme, and benchmarks it on two long-term vineyard datasets using both low-cost and conventional LiDAR sensors.

What This Paper Is About

Place recognition for mobile robots is hard in vineyards because the environment is unstructured, self-similar, and seasonally changing, so the geometry of vine rows offers few distinctive features to match against a stored map. Existing agricultural LPR methods underperform urban ones — the paper cites SPVSoAP3D reaching 76.30% Recall@1 in the most favorable agricultural case, versus 95.80% Recall@1 for the urban method MinkUNeXt — and agricultural localization remains largely unexplored compared with classification and segmentation. The goal is a lightweight network that runs in real time on low-cost, sparse LiDAR (such as a Livox) while still producing descriptors robust enough to recognize a place across seasons.

Key Contributions

1. MinkUNeXt-VINE, a lightweight LPR network for agricultural robotics that uses a Matryoshka Representation Learning multi-loss strategy plus architectural simplification to produce flexible low-dimensional descriptors for real-time use. The authors state this is the first contribution applying an MRL approach to optimize a LPR method.
2. An architectural trade-off analysis examining the interaction between MRL, reduced layer complexity, and descriptor dimensionality, including performance with a low-cost, low-resolution Livox LiDAR sensor.
3. Extensive cross-season validation of the method across the months and seasons making up the vineyard phenological cycle.
4. A benchmark of LPR approaches on recent long-term vineyard datasets, comparing the impact of sensors of differing cost and resolution, specifically Livox and Velodyne.

Main Findings

• Compact descriptors do not hurt performance in vineyards: Testing the original MinkUNeXt architecture at output dimensions of 512, 256, 192, 128, 64, 32, and 16 on Livox data from TEMPO-VINE, the paper reports that a more compact descriptor can match or exceed larger ones, because these unstructured scenes lack enough salient features to justify a high-dimensional descriptor. As one example, moving from 256 to 64 dimensions is reported to increase Recall@1 by +4% when evaluated in May (256: 20.61; 64: 24.73).
• Very low dimensionality degrades results: Dimensions of 32 or 16 begin to have a detrimental effect, so 192 was selected as the final descriptor size, chosen because performance was not compromised and it reported the greater Recall@1 in two of the three evaluations.
• Normalization of the input is the single largest design effect: With quantization size 0.01 and no normalization, the baseline reaches 31.99% Recall@1% and 6.13% Recall@1. Increasing the quantization size to 0.1 improves recall by +40%. Applying the proposed normalization to the point cloud raises Recall@1% above 90%.
• TSAP outperforms a simple contrastive loss: Replacing TSAP with Contrastive Loss drops results sharply (48.65% Recall@1% and 10.17% Recall@1, versus 95.59% and 58.82% for TSAP with the same settings). The authors attribute this to TSAP acting as a differentiable surrogate for Average Precision and optimizing global ranking, while triplet-style losses struggle with the perceptual aliasing of semi-structured vineyards.
• MRL beats other multi-loss combinations: Of the tested multi-loss setups (IDs 5–7) and MRL setups (IDs 8–9), the best configuration was MRL with the intermediate descriptor (ID 9): 97.30% Recall@1% and 67.77% Recall@1. The authors report MRL as a +5% improvement in Recall@1 relative to the comparable non-MRL multi-loss configuration.
• Extracting the descriptor earlier in the network helps: Using the output of the second transposed convolution as the final layer improved Recall@1 by almost +6% for the MRL approach. Compared with the original backbone setting (85.54% Recall@1%, 49.63% Recall@1), the final proposal improved by around +12% in Recall@1% and +18% in Recall@1.
• Seasonal generalization is uneven: In the seasonal study with "one in, one out" sampling, training in February gave 85.98% Recall@1% / 63.26% Recall@1 when tested on February, but fell to 58.76% / 23.12% when tested on May and 61.52% / 21.91% on the last June campaign. Training in May produced its lowest in-season result (May→May: 67.77% Recall@1%, 28.54% Recall@1) while generalizing relatively better to February (78.04% / 29.89%). The September-trained results in the paper's final table are not fully reported in the provided content.
• Two long-term vineyard datasets are used: The Bacchus Long-Term (BLT) dataset (Ktima Gerovassiliou, Greece, and Riseholme, UK; Ouster OS1-16 3D LiDAR) and TEMPO-VINE (Agliè, northern Italy; Velodyne VLP-16 and Livox MID360). The Velodyne point cloud contains approximately 10,000 more points per frame than the Livox counterpart.

Methodology in Plain English

The team starts from MinkUNeXt, an urban LiDAR place recognition network that uses only Minkowski convolutions in a U-Net-style encoder-decoder with skip connections. They make three changes. First, they preprocess raw scans: points with zero x, y, z values are filtered out, points beyond 60 meters are removed, and each point is normalized into a [-1, 1] range by scaling with a fixed factor S = 60 and subtracting the point cloud centroid. Second, they prune the network: following prior insight that descriptors taken after the second transpose convolution give minimized mean positional and orientation errors, they cut the next five layers and the final fully connected layer, place a GeM pooling aggregator right after that second transpose convolution, and read out a 192-dimensional descriptor instead of the original 512-dimensional one. Third, they change the training objective from a single TSAP loss to Matryoshka Representation Learning, which sums TSAP losses computed on nested prefixes of the descriptor; they use the dimensions [64, 128, 192] with weights [1.0, 0.5, 0.25], deliberately weighting lower dimensions more heavily for efficiency.

Evaluation uses Recall@1 and Recall@1%. For the benchmark section the trajectory is split into zones following the PointNetVLAD protocol, using a radius of 13 meters and three test reference points for BLT, and 23 meters with four test reference points for TEMPO-VINE. The ablation studies instead use only the Livox data from TEMPO-VINE, with either "run2" as the test set (40 to 50 percent of examples, depending on whether "run3" exists in the campaign) or a "one in, one out" 50/50 split. All three sensors record at 10 Hz, and scans are sampled by distance: every half meter for BLT and every meter for TEMPO-VINE. Ablation experiments ran on an Nvidia A30 GPU, with training set to 200 epochs — half of what the original MinkUNeXt paper proposes — because these datasets are considerably smaller.

Why This Matters

Research impact: The paper pushes back on the assumption that high-dimensional descriptors are needed for robust place recognition, showing that in feature-poor, repetitive agricultural scenes a compact descriptor can be equal or better. It is also presented as the first use of Matryoshka Representation Learning for LiDAR place recognition, and it provides a benchmark on the two long-term vineyard 3D LiDAR datasets that exist, giving the field a reproducible baseline. The code is publicly available at a GitHub repository linked in the paper.

Real-world applications:

• Autonomous navigation of agricultural vehicles that must localize without GPS.
• Loop closure detection and SLAM in field robotics, where a match to a previously visited place corrects drift.
• Precision agriculture and field/crop monitoring, where knowing exactly where the robot is enables per-plant data collection.
• Phenotyping, where repeated measurements of the same vines across a season require reliable re-identification of location.

Industry relevance: Determining position from a low-cost, sparse Livox sensor rather than an expensive high-resolution Velodyne matters commercially, because sensor cost is a major driver of agricultural robot pricing. The low dimensionality (192 instead of 512) also reduces memory footprint and search time, which supports real-time operation on field hardware.

Future Directions

• The cross-season results show a large drop from in-season to cross-season accuracy (for example February-trained models reaching 58.76% Recall@1% / 23.12% Recall@1 in May), so improving robustness to the vineyard phenological cycle and to the appearance change between bare winter vines and dense summer vegetation is the clearest open problem.
• The paper reports results for individual months and for season packages of three months each; how larger multi-season training sets compare to single-season training, and whether the September-trained models behave differently from the winter-trained ones, is only partially shown in the provided content.
• Extending the evaluation beyond the two available long-term vineyard datasets, and testing whether the MRL plus pruning recipe transfers to other crops and other unstructured agricultural settings, is a natural next step.
• The paper's comparisons include different LiDAR price and resolution tiers (Livox versus Velodyne, and the Ouster OS1-16 on BLT); further work could quantify how far the descriptor can be compressed before the 32- and 16-dimensional degradation observed in the dimensionality study becomes limiting in harder conditions.

Target Audience

Robotics and computer vision researchers working on place recognition, SLAM, and loop closure detection; engineers building autonomous agricultural vehicles or field robots that must operate without GPS; and practitioners interested in efficient descriptor design, Matryoshka Representation Learning, or sparse-convolution backbones. It will also be useful to agri-tech developers choosing between low-cost and high-resolution LiDAR sensors, and to researchers looking for established benchmark datasets and splits for vineyard localization.

Authors’ 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.