Temperate forests are easy. The canopy is structured — dominant species, regular spacing, visible gaps between crowns. A segmentation algorithm can find individual trees because the trees cooperate by being distinct.
Tropical forests refuse this courtesy. Canopies overlap in three dimensions. Species diversity means no two adjacent trees look alike. Lianas and epiphytes blur the boundary between one crown and another. The density that makes tropical forests critical for carbon storage — canopy trees store disproportionate biomass — is exactly what makes them impossible to segment with off-the-shelf methods.
SelvaMask (arXiv:2602.02426) provides 8,861 manually delineated tree crowns across three Neotropical sites in Panama, Brazil, and Ecuador, at 1.3–3.5 cm per pixel from drone imagery. It's the largest tropical tree crown segmentation dataset, roughly twice existing benchmarks. The key methodological choice: a modular detection-segmentation pipeline that adapts vision foundation models using domain-specific detection prompting, rather than training end-to-end on limited data.
The structural insight is in the inter-annotator agreement evaluation. When even human experts disagree on where one crown ends and another begins, the benchmark is measuring the fundamental ambiguity of the problem, not just the model's accuracy. The “ground truth” in tropical canopy segmentation is contested by definition. The dataset doesn't resolve this — it quantifies it. That's more honest and more useful than pretending the labels are clean.