State where, when and for which targets performance was tested; keep untested settings visible.
Avoid near-duplicate leakage
Adjacent frames may share nearly identical animals and background. Splitting them across training and testing can make performance appear stronger than it is. Consider sortie, date, location and individual identity when partitioning data. Reserve an independent setting for final evaluation. No percentage split automatically solves dependence; explain how independence was protected.
Report decision-relevant errors
Precision concerns the correctness of detections; recall concerns the labelled targets found. Add count bias, plot-level error, species and size breakdowns, and review time to detection metrics. Rare-species decisions can be sensitive to very few misses. Check confidence calibration instead of assuming that a model score is an empirical probability.
Treat domain shift as an operating condition
Canopy, snow, season, thermal background, height, optics and compression can change model behaviour. WAID includes cross-dataset evaluation, but benchmark success is not a field guarantee. Build a local validation set and document unsupported conditions. After adaptation, retest earlier conditions as well as the new site to detect regressions.
Audit the reference labels
Human annotations contain misses and disagreements too. Use independent review and adjudication for difficult examples, preserving genuine uncertainty. Provide original images, temporal context and paired sensor views. Approve model updates based on consequential errors and reproducibility, rather than one average metric.
Evidence & further reading
Check the original study conditions, species and location before using a result in a plan. Do not transfer findings universally across species and settings.