Abstract

Fine-grained localisation of plumage regions is a prerequisite for computational analyses of avian colouration, patterning and visual traits in ecological and evolutionary research. Progress is limited by the scarcity of image resources with annotations aligned to biologically meaningful anatomical units: existing avian benchmarks provide either landmark points or coarse part categories that do not capture ornithologically defined plumage regions. We present a curated dataset of 4,705 bird images annotated for nine plumage regions: head, throat, breast, belly, vent, back, coverts, remiges and tail. Spanning 39 avian orders and 222 families, the dataset provides a taxonomically broad resource for fine-grained avian image analysis. The dataset was built through an iterative model-assisted annotation workflow, in which model predictions were reviewed and corrected rather than drawn from scratch, improving the efficiency of region-level annotation. We benchmark classical segmentation architectures, SAM-based models and self-supervised foundation-model encoders on this task. A frozen DINOv3 encoder with a lightweight decoder achieved the highest performance on the held-out test set, reaching 84.01% mean Intersection over Union while requiring substantially less memory than end-to-end fine-tuning. The model generalised to external avian benchmarks, including the bird subset of PartImageNet and CUB-200-2011, and achieved competitive performance on the full PartImageNet part-segmentation benchmark, which includes diverse animal taxa. We provide a modular detect-track-segment pipeline as a proof-of-concept extension to video data. Together, these results show that anatomically grounded avian annotations can serve both as a resource for plumage phenotyping and as a benchmark for efficient, transferable biological part segmentation.