TAT-VPR: Ternary Adaptive Transformer for Dynamic and Efficient Visual Place Recognition

2025-05-22

AI Summary

To address the trade-off between accuracy and efficiency in visual SLAM loop closure detection under dynamic scenes, this paper proposes a ternary adaptive Transformer architecture for dynamically adjustable visual place recognition. Our method introduces two key innovations: (1) a novel dynamic computation control mechanism that jointly integrates ternary-weight quantization and a learnable sparse activation gate, enabling runtime on-demand adjustment of computational load; and (2) a two-stage knowledge distillation pipeline that preserves descriptor discriminability under ultra-low-bit constraints. Experiments demonstrate that our approach achieves zero-loss Recall@1 while reducing peak computational cost by up to 40%, significantly outperforming existing lightweight methods. Furthermore, it has been successfully deployed on micro aerial vehicles and embedded SLAM systems, attaining state-of-the-art localization accuracy.

Technology Category

• Computer Vision: Learning & Optimization for CV
• Machine Learning: Hardware-aware ML
• Search and Optimization: Learning to Search

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

TAT-VPR is a ternary-quantized transformer that brings dynamic accuracy-efficiency trade-offs to visual SLAM loop-closure. By fusing ternary weights with a learned activation-sparsity gate, the model can control computation by up to 40% at run-time without degrading performance (Recall@1). The proposed two-stage distillation pipeline preserves descriptor quality, letting it run on micro-UAV and embedded SLAM stacks while matching state-of-the-art localization accuracy.

Problem

• Dynamic accuracy-efficiency trade-offs in visual SLAM loop-closure
• Run-time computation control without performance degradation
• Preserving descriptor quality for micro-UAV and embedded SLAM

Innovation

• Ternary-quantized transformer for dynamic efficiency
• Learned activation-sparsity gate controls computation
• Two-stage distillation preserves descriptor quality