Faster or Stronger: Towards Flexible Visual Place Recognition via Weighted Aggregation and Token Pruning
The paper presents advancements in Visual Place Recognition (VPR) through a new method called Weighted Aggregated Descriptor (WeiAD). This method improves the aggregation of patch tokens by assigning weights to clusters, enhancing the accuracy of global representations. Additionally, the authors introduce WeiToP, a token pruning framework that optimizes feature extraction costs while maintaining performance.
- ▪Visual Place Recognition aims to match query images to reference images in large-scale databases.
- ▪The proposed WeiAD method enhances VPR performance by weighting clusters during aggregation.
- ▪WeiToP allows for flexible token pruning at inference time, improving efficiency without additional training.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computer Vision and Pattern Recognition arXiv:2605.20551 (cs) [Submitted on 19 May 2026] Title:Faster or Stronger: Towards Flexible Visual Place Recognition via Weighted Aggregation and Token Pruning Authors:Zichao Zeng, June Moh Goo, Junwei Zheng, Weijia Fan, Jiaming Zhang, Rainer Stiefelhagen, Jan Boehm View a PDF of the paper titled Faster or Stronger: Towards Flexible Visual Place Recognition via Weighted Aggregation and Token Pruning, by Zichao Zeng and 6 other authors View PDF HTML (experimental) Abstract:Visual Place Recognition (VPR) aims to match a query image to reference images of the same place in a large-scale database.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
