Cross-dataset few-shot deep learning for landmark detection in biological morphometrics
Year: 2026
Type: article
Abstract: Geometric morphometrics relies on biologically homologous landmarks to quantify shape variation, yet acquiring sufficient landmark annotations remains a major obstacle to automated analysis, particularly for small, rare, or newly collected biological datasets. We propose an annotation-efficient deep learning framework for biological landmark detection that separates transferable visual representation learning from target-specific landmark adaptation. The framework first learns a general visual representation through masked autoencoding and then adapts it to a new landmark configuration using few-shot heatmap regression. This design avoids supervised episodic meta-training and allows the same pretrained encoder and adaptation procedure to be reused across datasets with different species, anatomical structures, imaging modalities, and landmark configurations. Importantly, the framework also supports target-excluded transfer, in which representation learning is performed without access to the target dataset before adaptation using only a small annotated subset. Evaluation on heterogeneous biological landmark datasets shows that the proposed framework can achieve reliable localization with limited supervision while maintaining a common architecture and training strategy across substantially different morphometric tasks. By reducing both annotation requirements and dataset-specific model engineering, the proposed framework offers a scalable route for applying deep landmark detection to small and heterogeneous biological collections.
Source: Ecological Informatics
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,770
OpenAlex ID: W7223151180