通过概率重加权，让匹配模型在稀疏与密集间自由切换。
Transferring between sparse and dense matching via probabilistic reweighting
- 用特征检测概率重调注意力与匹配分数，实现跨稀疏度适配。
- 重加权后稀疏LoFTR性能接近有检测器的模型，精度提升显著。
- 无需修改参数，适用于SuperGlue、LightGlue等主流匹配网络。
基于检测器和无检测器的匹配方法仅适用于特定稀疏范围。为提升现有匹配器的适应性，本文提出一种新型概率重加权方法。该方法适用于基于Transformer的匹配网络，可在不改变网络参数的前提下适应不同稀疏水平。重加权通过调整特征的检测概率来优化注意力权重与匹配分数，并证明重加权匹配网络是基于检测器匹配网络的渐近极限。此外，我们基于重加权设计了稀疏训练与剪枝流程，应用于无检测器网络。对SuperGlue、LightGlue和LoFTR的重加权版本在多种稀疏度下进行评估。实验表明，该方法提升了基于检测器匹配器在密集特征上的姿态精度；重加权稀疏LoFTR性能可媲美基于检测器的匹配器，展现出良好的精度与计算复杂度平衡能力。
原文摘要 · Abstract (English)
Detector-based and detector-free matchers are only applicable within their respective sparsity ranges. To improve adaptability of existing matchers, this paper introduces a novel probabilistic reweighting method. Our method is applicable to Transformer-based matching networks and adapts them to different sparsity levels without altering network parameters. The reweighting approach adjusts attention weights and matching scores using detection probabilities of features. And we prove that the reweighted matching network is the asymptotic limit of detector-based matching network. Furthermore, we propose a sparse training and pruning pipeline for detector-free networks based on reweighting. Reweighted versions of SuperGlue, LightGlue, and LoFTR are implemented and evaluated across different levels of sparsity. Experiments show that the reweighting method improves pose accuracy of detector-based matchers on dense features. And the performance of reweighted sparse LoFTR is comparable to detector-based matchers, demonstrating good flexibility in balancing accuracy and computational complexity.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。