Measurement-guided, training-free Fourier correction: a cost-efficient alternative to generative adaptation for cold-start construction-scene segmentation

<title>Abstract</title> <p>Deploying pixel-level scene understanding at a new construction site is a cold-start problem: only a handful of labeled real frames are available, and the rare but safety- and task-critical classes---workers and dump piles---are precisely the hardest to learn. Labeled synthetic data can fill this gap, but the sim-to-real appearance mismatch limits its value, and the usual remedies, unsupervised domain adaptation and generative image translation, assume abundant unlabeled target data and compute that a cold-start site does not have. We introduce a measurement-guided correction protocol that (i) decomposes the low-frequency Fourier amplitude into global colour/exposure and texture-related components, (ii) quantifies each mismatch with sliced-Wasserstein distances, and (iii) verifies which component drives the downstream error before applying a single training-free pass of Fourier domain adaptation. On an excavator-worksite dataset with only 1% real labels, the protocol localizes the correctable mismatch to insufficient texture variation in simulated soil; correcting the low-frequency amplitude raises rare-class Dump IoU from 0.465 to 0.565 (\((n=6)\) seeds, \((p&lt;0.001)\)) by suppressing false-positive predictions on rough real regions. The training-free correction significantly outperforms generative translation with ControlNet, is statistically indistinguishable from far more expensive per-image diffusion restyling, and surpasses classical colour and augmentation corrections, at a fraction of their data and compute budget. The protocol is released as a reusable, measurement-first recipe for cold-start sim-to-real segmentation.</p>