Matching Support and Computational Cost of SuperPoint Front Ends for Underwater Images

Underwater image processing trades correspondence availability against computational cost. We examine this trade-off using SuperPoint configurations with raw, contrast-enhanced, and joint image processing under a common matching pipeline. On 200 AQUALOC harbor pairs, dual-branch extraction yielded 166.52 essential-matrix inliers on average, compared with 131.32 for raw-image SuperPoint, and exceeded both single branches on 196 pairs. A sequence-held-out allocation policy yielded 155.49 inliers, compared with 153.84 for random reassignment with the same candidate totals. After removing output-equivalent redundant computation, direct measurements on all 200 pairs showed that the frozen policy retained 93.4% of dual-branch mean support with 31.1% lower mean processing time. Training-only cost recalibration and simple allocation baselines characterize the dependence of this operating point on implementation and decision rule. Matched-keypoint controls at four budgets did not establish an equal-count advantage under common confidence-ranked truncation. Archaeological imagery and frozen SubPipe transfer distinguish candidate support from portable allocation, while short-segment visual odometry did not establish an overall tracking improvement. The study quantifies the support supplied by additional processing and the cost of allocating it, providing an evaluated basis for downstream-aware allocation.