TeTRA-VPR: A Ternary Transformer Approach for Compact Visual Place Recognition

2025-03-04

AI Summary

For resource-constrained platforms such as UAVs and mobile robots, visual place recognition (VPR) faces challenges including large model size, high inference latency, and difficulty balancing accuracy and efficiency. This paper proposes a lightweight and efficient VPR method: the first joint compression paradigm integrating a progressive ternarized Vision Transformer (ViT) backbone with a binarized embedding layer, coupled with a representation-preserving progressive distillation strategy. This enables ultra-low-bit models (2-bit backbone + 1-bit embedding) to surpass full-precision CNN baselines in accuracy. Experiments show that, compared to state-of-the-art efficient baselines, our method reduces memory footprint by 69% and inference latency by 35%, while maintaining or slightly improving Recall@1—achieving, for the first time, high-accuracy, low-overhead VPR deployment on edge devices.

Technology Category

• Computer Vision: Representation Learning for Vision
• Machine Learning: Learning on the Edge & Model Compression
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Systems and Infrastructure for Web, Mobile and WoT: Virtualization and resource management in Web systems and infrastructures
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs

Abstract

Visual Place Recognition (VPR) localizes a query image by matching it against a database of geo-tagged reference images, making it essential for navigation and mapping in robotics. Although Vision Transformer (ViT) solutions deliver high accuracy, their large models often exceed the memory and compute budgets of resource-constrained platforms such as drones and mobile robots. To address this issue, we propose TeTRA, a ternary transformer approach that progressively quantizes the ViT backbone to 2-bit precision and binarizes its final embedding layer, offering substantial reductions in model size and latency. A carefully designed progressive distillation strategy preserves the representational power of a full-precision teacher, allowing TeTRA to retain or even surpass the accuracy of uncompressed convolutional counterparts, despite using fewer resources. Experiments on standard VPR benchmarks demonstrate that TeTRA reduces memory consumption by up to 69% compared to efficient baselines, while lowering inference latency by 35%, with either no loss or a slight improvement in recall@1. These gains enable high-accuracy VPR on power-constrained, memory-limited robotic platforms, making TeTRA an appealing solution for real-world deployment.

Problem

• Reduces memory and compute requirements for Visual Place Recognition.
• Enables high-accuracy VPR on resource-constrained robotic platforms.
• Achieves significant model size and latency reduction without accuracy loss.

Innovation

• Ternary transformer reduces model size
• Progressive distillation preserves accuracy
• 2-bit quantization lowers memory usage