Answers are grounded in this paper and cite their sources.
Algorithm evaluation demonstrates enhanced semi-dense image matching accuracy and efficiency across diverse photogrammetric benchmarks, indicating the value of front-end noisy pose integration.
In operational photogrammetry and remote sensing, image pairs are often accompanied by approximate pose information from GNSS/IMU, coarse registration, SLAM, or upstream block adjustment. Yet most modern matching pipelines still treat matching largely as a blind search problem, spending computation on geometrically implausible regions and leaving more room for outliers. Directly exploiting pose priors is non-trivial because these priors are themselves noisy. To address this issue, we propose a Joint Orientation and Pose-Aware Attention Network (JOPA-Net), a detector-free framework that injects pose priors into the matching process to improve robust semi-dense image matching. JOPA-Net combines three tightly coupled components: a MambaVision-based multi-scale feature pyramid, a geometry-guided routing attention module that concentrates cross-view interaction within epipolar-feasible bands, and a differentiable confidence-weighted pose refinement module that feeds updated geometry back to subsequent matching stages across scales. To accommodate prior uncertainty, the epipolar band is adaptively widened through covariance propagation and uncertainty-aware bandwidth control. We evaluate the method across four representative scenarios, including close-range photogrammetry as well as vehicle, airborne, and satellite imagery. For fair comparison, competing methods that do not use pose priors are additionally equipped with the same prior-derived post-hoc geometric filtering. Results show consistent gains across all evaluations. On the challenging ScanNet benchmark, JOPA-Net reaches semi-dense matching accuracy comparable to or better than representative dense methods while remaining markedly more efficient. These results indicate that integrating noisy pose priors into the front end of matching is an effective way to improve robustness and efficiency in practical photogrammetric pipelines.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: