How-to
Where to get LiDAR-verified no-change examples for change detection models
Every change-detection method needs to be tested on terrain that genuinely didn't change — hard negatives, not just positives. Real ones are rare to find pre-labeled.
3 steps
- Use real stable terrain, not simulated — Real LiDAR noise and real co-registration offset behave differently than synthetic data — a method that only sees clean simulated negatives can be miscalibrated on real data.
- Keep the rejected candidates too — A dataset that only shows you the negatives that passed its own quality filter is hiding its failure mode — the honest version publishes both.
- Check the verdict was computed by the same method you're validating against — A negative control is only useful if it was produced by a real, published, generalized method — not a one-off script for this dataset alone.