Search

Local feature matching constitutes the cornerstone of multiple computer vision applications (e.g., 3D reconstruction and long-term visual localization), and has been successfully resolved by detector-free methods. To further improve the matching performance, more recent research has focused on designing sophisticated architectures but endures additional computational overhead. In this study, with a different perspective from previous studies, we aim to develop a deep and compact matching network to improve performance while reducing computing cost. The key insight is that a local feature matcher with deep layers can capture more human-intuitive and simpler-to-match features. To this end, we propose DeepMatcher, a deep transformer-based network that tackles the inherent obstacles of not being able to build a deep local feature matcher with current methods. DeepMatcher consists of: (1) a local feature extractor (LFE), (2) a feature-transition module (FTM), (3) a slimming transformer (SlimFormer), (4) a coarse matches module (CMM), and (5) a fine matches module (FMM). The LFE is utilized to generate dense keypoints with enriched features from the images. We then introduce the FTM to ensure a smooth transition of feature scopes from LFE to the subsequent SlimFormer because of their different receptive fields. Subsequently, we develop SlimFormer dedicated to DeepMatcher, which leverages vector-based attention to model the relevance among all keypoints, enabling the network to construct a deep Transformer architecture with less computational cost. Relative position encoding is applied to each SlimFormer to explicitly disclose the relative distance information, thereby improving the representation of the keypoints. A layer-scale strategy is also employed in each SlimFormer to enable the network to adaptively assimilate message exchange, thus endowing it to simulate human behavior, in which humans can acquire different matching cues each time they scan an image pair. By interleaving the self- and cross-SlimFormers multiple times, DeepMatcher can easily establish pixel-wise dense matches at the coarse level using the CMM. Finally, we consider match refinement as a combination of classification and regression problems and design an FMM to predict confidence and offset concurrently, thus generating robust and accurate matches. Compared with our baseline LoFTR in indoor/outdoor pose estimation, DeepMatcher surpasses it by 3.32%/2.91% in AUC@5 ∘. Besides, DeepMatcher and DeepMatcher-L significantly reduce computational cost and only consume 77.89% and 92.46% GFLOPs of LoFTR. Large DeepMatcher considerably outperforms state-of-the-art methods on several benchmarks, including outdoor pose estimation (MegaDepth dataset), indoor pose estimation (ScanNet dataset), homography estimation (HPatches dataset), and image matching (HPatches dataset), demonstrating the superior matching capability of a deep local feature matcher. • We propose DeepMatcher, a deep Transformer-based network for local feature matching. • We propose SlimFormer to enable DeepMatcher to be extended into dozen layers. • We propose FTM to ensure a smooth transition from CNN to SlimFormer. • We propose FMM to optimize coarse matches, deriving robust and accurate matches. [ABSTRACT FROM AUTHOR]

Local feature matching involves the task of establishing the pixel-wise correspondences between a pair of images. As an integral component of plentiful computer vision applications (e.g., visual localization), this task has been successfully performed using Transformer-based methods. However, these methods typically extract numerous keypoints from sparse texture regions to construct a densely-connected graph neural network (GNN) for long-range feature aggregation, which inevitably triggers redundant message exchange and hampers the learning process. Furthermore, they employ transformer encoder layers that consider images as 1D sequences, leaving them incapable of extracting multiscale local structural information from the images, which is critical for establishing correspondence in image pairs with significantly scales shifts. In this study, we develop FMAP, an innovative detector-free approach that enables accurate local feature matching. For the first issue, FMAP employs an anchor points feature aggregation module (APAM) that captures representative keypoints and discards the extraneous keypoints to build a sparsified GNN for compact yet clean message exchange, with the key insight that the keypoints containing abundant visual information are distinguishable from their neighbors. For the second issue, FMAP proposes a global–local multiscale perception module (GMPM), which incorporates abundant multiscale local context information into global feature representation by employing multiple depth-wise convolutions with varying kernel sizes, thereby generating discriminative features that are robust to scale shifts. In addition, the depth-wise convolutions are utilized in the feed-forward network of the GMPM to further fuse the global context information and local feature representation. Extensive experiments on several standard benchmarks demonstrate that the proposed FMAP method significantly outperforms state-of-the-art methods. Compared to the cutting-edge methods MatchFormer, QuadTree, and TopicFM in relative pose estimation task, FMAP surpasses them by 2.27%, 0.58%, and 1.08% in terms of AUC@5°. Besides, FMAP noticeably outperforms the baseline LoFTR by (2. 38 % , 1. 89 % , 1. 45 %) in terms of AUC@(5°, 10°, 20°). Moreover, we integrate FMAP into an official visual localization framework and conduct a visual localization experiment, with the results showing that FMAP exceeds LoFTR by 2.3% in terms of AP. • We propose a Transformer-based method that achieves excellent matching accuracy. • We constructs sparsified GNN to alleviate redundant message exchange. • We integrate global context and multi-scale local features to boost robustness. [ABSTRACT FROM AUTHOR]