Flower-Counting Pipeline Scores Perfectly on Its Own Simulator but Falters Under Light and Occlusion

A pipeline that counts open blooms and closed buds in crop photos reported 99.7% recall on its own simulated scenes. The author argues this score is misleading because the simulator and detector were tuned to each other, and the scoring relied on totals rather than per-object matching. To test robustness, the author applied lighting changes such as gamma shifts, shadows, and color casts, along with leaf occlusion and clustered flowers, to the unmodified detector. Under strict per-object scoring, the untouched scenes still gave perfect results, which the author attributes to the co-tuned design. The article then evaluates how the detector degrades under these conditions and whether two simple preprocessing and splitting remedies help.