EdgeLPR: On the Deep Neural Network trade-off between Precision and Performance in LiDAR Place Recognition

Technical Summary: EdgeLPR

Problem Statement
Long-term autonomous navigation relies heavily on place recognition to achieve loop closure and maintain consistent mapping. While deep learning has significantly improved the accuracy of extracting discriminative features from sensory data, deploying these models on resource-constrained EdgeAI platforms remains a critical bottleneck. Traditional LiDAR-based approaches often require storing and comparing large point clouds, leading to prohibitive memory and computational demands. Furthermore, the concurrent deployment of multiple deep networks (for state estimation, mapping, detection, etc.) exacerbates these constraints. There is a lack of systematic analysis regarding how lightweight architectures and numerical precision trade-offs affect performance in LiDAR place recognition, particularly when transitioning from high-precision floating-point to low-precision integer inference.

Methodology
The paper proposes EdgeLPR, a framework for efficient LiDAR-based place recognition that leverages Bird's Eye View (BEV) representations to enable the use of lightweight, image-based convolutional neural networks (CNNs).

• Representation: Instead of processing raw 3D point clouds, the system projects LiDAR scans into 256x256 BEV images. This reduces input dimensionality and allows the use of efficient 2D convolutional backbones.
• Architecture: The framework employs a unified descriptor scheme. It utilizes three representative lightweight backbones: MobileNetV3, ShuffleNetV2, and ResNet18. Unlike prior works that rely on complex aggregation heads (e.g., NetVLAD), these models extract features, apply adaptive global pooling, and pass the result through a shallow Fully Connected (FC) layer to generate a compact global descriptor. This design isolates the contribution of the backbone architecture and ensures a consistent descriptor size for fair comparison.
• Quantization Strategy: To evaluate the trade-off between efficiency and accuracy, the authors benchmark the models under three precision regimes: FP32 (full precision), FP16 (half-precision), and INT8 (8-bit integer). The INT8 evaluation utilizes a static post-training quantization (PTQ) strategy, where scaling factors and zero-points are determined via a calibration phase to minimize reconstruction error on a representative subset of data.
• Training and Evaluation: The models are trained on the KITTI dataset (sequence '00') using a triplet loss objective with random in-plane rotation augmentation. Evaluation is conducted on two datasets:
  • KITTI (sequences 02, 05, 06): For intra-sequence, short-term viewpoint robustness.
  • NCLT: For inter-sequence, long-term seasonal and appearance variations.
    Performance is measured using Recall@k (k=1, 5, 10, 20), F1-score, PR AUC, and Mean Reciprocal Rank (MRR).

Key Contributions

1. Novel Evaluation Framework: The design and validation of a controlled, reproducible framework for evaluating lightweight neural networks under varying quantization levels specifically for place recognition.
2. Comprehensive Quantitative Assessment: A systematic analysis of state-of-the-art lightweight architectures (MobileNetV3, ShuffleNetV2, ResNet18) across different precisions and environmental complexities (short-term vs. long-term).
3. Empirical Toolbox: The provision of data-driven insights to support hardware and performance-informed design choices for next-generation robotic perception systems, specifically addressing the gap in understanding how input representation and model complexity interact with resource constraints.

Results
The experiments reveal distinct trade-offs between accuracy, robustness, and efficiency:

• FP16 vs. FP32: Across all architectures, FP16 yields results nearly identical to FP32, offering a "free" gain in memory efficiency and computational speed without compromising accuracy.
• INT8 Impact: The degradation caused by INT8 quantization is highly architecture-dependent.
  • MobileNetV3: Shows the highest accuracy in floating-point but suffers the most significant performance drop under INT8 (e.g., Recall@1 on KITTI drops from ~0.90 to ~0.75).
  • ShuffleNetV2: Demonstrates the most favorable balance, maintaining competitive performance with minimal degradation under INT8 while offering the smallest memory footprint (~2.17 MB in INT8).
  • ResNet18: Exhibits the strongest robustness to INT8 quantization and long-term appearance changes (NCLT dataset), with negligible performance degradation. However, its large memory footprint (~11.15 MB in INT8) limits its suitability for strictly constrained platforms.
• Inference Behavior: Interestingly, parameter efficiency does not always translate to hardware efficiency. For instance, ShuffleNetV2 showed limited throughput scalability under INT8 CPU execution compared to its floating-point performance, suggesting that architectural efficiency in terms of parameters does not guarantee efficient hardware utilization in all contexts.

Significance and Claims
The paper claims to provide a principled foundation for developing resource-efficient, AI-driven place recognition pipelines. It argues that the optimal architecture is not universal but depends on specific target constraints:

• If maximum accuracy is the priority and moderate resources are available, MobileNetV3 in FP16 is a strong candidate.
• For strict memory and power budgets, ShuffleNetV2 with INT8 offers the most efficient solution, achieving competitive performance with a ~4x reduction in memory compared to FP32.
• ResNet18 is highlighted for its robustness to quantization and temporal variations, though its size limits low-power applicability.

The authors emphasize that robustness to quantization alone is insufficient; it must be paired with architectural efficiency. The work concludes that co-designing architecture, descriptor dimensionality, and numerical precision is essential for scalable, long-term deployment on low-power robotic platforms. Future work is suggested to explore joint optimization strategies, such as quantization-aware training and descriptor dimensionality reduction.