The Resurgence of Local Feature Matching: Why Data Scale, Not Architecture, Explains the Detector-Free Revolution

Introduction

For the past several years, the computer vision community has operated under a particular narrative regarding image matching: detector-based local feature methods were becoming obsolete, rendered unnecessary by the rise of detector-free dense matching and feed-forward reconstruction models. Approaches such as LoFTR, RoMa, MASt3R, and VGGT appeared to demonstrate that sparse keypoint detection was a fundamental bottleneck, unable to handle textureless regions, extreme viewpoint changes, or challenging illumination conditions. The conventional wisdom held that the future belonged to dense correspondence prediction or direct 3D reconstruction from pixels, with local feature matching relegated to legacy systems requiring computational efficiency.

The paper "LoMa: Local Feature Matching Revisited" by Nordström et al. presents a compelling challenge to this assumption. Through systematic scaling of training data, model capacity, and compute, the authors demonstrate that local feature matching was not architecturally inferior; it was simply undertrained. Their results suggest that the performance gap between sparse and dense methods reflected differences in training regimes rather than inherent limitations of keypoint-based approaches. This finding carries significant implications for how we evaluate architectural innovations and allocate computational resources in 3D vision research.

The Scaling Hypothesis and Training Regimes

The central thesis of LoMa rests on a straightforward but previously underexplored observation: while dense matching and feed-forward reconstruction methods have been trained on internet-scale datasets, local feature matching models have remained constrained to a few mid-sized datasets, primarily MegaDepth and ScanNet. This disparity in data scale created an unfair comparison that the field interpreted as architectural superiority.

Nordström et al. address this by introducing modern training recipes for local feature descriptors and matchers, combined with significantly larger and more diverse training mixtures. Their approach, termed LoMa, scales along two axes: data diversity (using MVS-generated ground truth from multiple sources) and model capacity (with three variants: LoMa-B, LoMa-L, and LoMa-G for Base, Large, and Gigantic). The results are substantial. On the challenging WxBS benchmark, LoMa achieves a +29.5 mAA improvement over the previous state-of-the-art ALIKED+LightGlue. On InLoc, the gain is +21.4 at the (1m, 10°) threshold, and on the IMC 2022 benchmark, the improvement reaches +12.4 mAA.

These gains are not incremental optimizations; they represent a qualitative shift in performance that places local feature methods competitive with, and in some cases superior to, dense matching approaches. The scaling laws observed in the paper suggest that local feature matching exhibits the same responsiveness to data scale and model capacity that has driven progress in other areas of computer vision, yet this dimension remained unexplored due to the field's premature conclusion that the architecture had reached its limits.

HardMatch and the Problem of Benchmark Saturation

Beyond the training contributions, LoMa introduces a critical methodological advance in evaluation through the HardMatch benchmark. The authors identify a significant problem in existing evaluation protocols: standard benchmarks like MegaDepth and ScanNet are derived from successful Structure-from-Motion reconstructions. This creates a selection bias toward image pairs that are already relatively easy to match, as failed reconstructions are discarded from the dataset.

As shown in the paper's analysis, modern state-of-the-art matchers already achieve high Percentage of Correct Keypoints (PCK) on the majority of pairs in these datasets, leading to benchmark saturation. When a large fraction of image pairs are already successfully matched by existing methods, further improvements become statistically difficult to measure, and the benchmarks lose their discriminative power. Worse, this saturation encourages overfitting to specific geometric verification settings or dataset artifacts rather than fundamental improvements in matching robustness.

HardMatch addresses this by providing 1000 manually annotated image pairs across 9 challenging categories, including extreme viewpoint changes, heavy occlusion, and significant illumination variation. Unlike SfM-derived ground truth, which cannot provide reliable correspondences for pairs that fail to reconstruct, HardMatch uses human annotation to establish ground truth even in cases where current algorithms fail completely. The benchmark reveals that feed-forward reconstruction methods largely fail on these hard cases, and even state-of-the-art dense matchers struggle significantly. In this challenging regime, LoMa achieves +18.6 mAA over ALIKED+LightGlue, demonstrating that the scaling of local feature matching specifically improves robustness on cases previously considered beyond the capability of sparse methods.

Reconsidering Architectural Trade-offs

The LoMa results force a reconsideration of the architectural debate between detector-based and detector-free methods. For years, the field has accepted a trade-off: dense methods offer superior accuracy and robustness at the cost of computational efficiency, while sparse methods offer speed suitable for large-scale SfM but with reduced accuracy on difficult pairs. If this trade-off was largely an artifact of training data scale, the implications for practical system design are profound.

Local feature matching offers distinct operational advantages that remain relevant for production systems. Sparse keypoint detection enables sublinear matching complexity with respect to image resolution, reduced memory footprints, and compatibility with decades of geometric verification and bundle adjustment research. If these advantages can be retained while achieving accuracy comparable to dense methods, the efficiency benefits for large-scale 3D reconstruction, visual localization, and real-time applications are substantial.

However, this does not imply that dense methods have no place in the ecosystem. Rather, LoMa suggests that the field may have been too hasty in declaring the fundamental superiority of one architectural family over another. The performance characteristics of different approaches may be more fluid than previously assumed, dependent on the specific data distributions and computational budgets employed during training.

Critical Analysis and Limitations

While the LoMa paper presents strong evidence for the scaling hypothesis, several nuances warrant consideration. First, the HardMatch benchmark, though necessary and well-designed, remains relatively small with 1000 pairs. Manual annotation is expensive and does not scale easily, potentially limiting the size of future training sets for the hardest cases. The community will need to develop semi-automated or verified automated methods to expand such benchmarks without prohibitive costs.

Second, local feature matching retains certain fundamental limitations that data scaling cannot entirely eliminate. In truly textureless regions, keypoint detection is inherently ill-posed; no amount of training data enables the detection of corners where none exist geometrically. While LoMa improves matching robustness for detected keypoints, it does not solve the coverage problem in homogeneous regions. Dense methods may retain an advantage in these specific scenarios, suggesting that hybrid approaches combining sparse and dense predictions may ultimately represent the optimal path forward.

Additionally, the computational costs of training LoMa-G, the gigantic variant, are substantial. The paper demonstrates that scaling benefits local feature matching, but it does not fully address whether these training costs are justified compared to simply scaling dense methods further. The inference speed advantages of local features must be weighed against the environmental and economic costs of training larger models, a consideration that grows more pressing as the field scales compute.

Conclusion and Future Directions

LoMa represents a methodological correction for the 3D vision community. By demonstrating that local feature matching responds robustly to data and compute scaling, Nordström et al. challenge the narrative that architectural innovation alone drives progress in image matching. Their work suggests that many of the perceived limitations of sparse methods were actually symptoms of insufficient training resources.

Looking forward, the field must address several open questions. Can the scaling laws observed for LoMa continue with orders of magnitude more data, or will local feature matching encounter inherent capacity constraints before matching the data efficiency of dense approaches? How can the community build larger versions of HardMatch to ensure benchmarks remain ahead of model capabilities? And crucially, can the efficiency advantages of sparse detection be combined with the representational capacity of dense features in unified architectures?

The most important contribution of LoMa may not be the specific model weights or even the benchmark, but the reminder that in empirical disciplines, we must isolate architectural variables from training variables before declaring the superiority of one approach over another. In the case of local feature matching, the obituary was written prematurely. The architecture remains competitive, efficient, and worthy of continued investment, provided we afford it the same computational resources given to its alternatives.