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

TL;DR

TAT-VPR introduces a ternary-quantized transformer that dynamically balances accuracy and efficiency in visual place recognition, enabling deployment on resource-constrained platforms without sacrificing localization performance.

Contribution

It proposes a novel ternary-quantized transformer with a learned activation-sparsity gate and a two-stage distillation pipeline for efficient and accurate visual SLAM loop-closure.

Findings

Reduces computation by up to 40% at run-time.

Maintains state-of-the-art localization accuracy.

Operates effectively on micro-UAV and embedded systems.

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.