Publikation
Computer Vision for Ecology
Robert Leist; Thiago Gouvea; Daniel Sonntag (Hrsg.)
European Conference on Computer Vision (ECCV-2026), Computer Vision for Ecology at the 19th European Conference on Computer Vision, located at ECCV-2026, September 8-12, Malmö, Sweden, Springer, 2026.
Zusammenfassung
Research such as marine ecology relies heavily on the analysis of corals and their morphologies from images, yet segmentation of instances remains a hard computer vision task: even foundation models such as SAM~3 typically require extensive annotation and retraining to perform well on them. We present Foveate, a training-free method that segments instances recursively leveraging frozen DINOv3~\cite{simeoniDINOv32025} embeddings, given as few as one exemplar mask of a target instance. Foveate mimics the human fovea by recursively redirecting the vision encoder's attention to candidate regions: starting from the whole image, it gates patches by their similarity to an exemplar bank, proposes tighter crops from the connected components of the foreground, and re-embeds each crop, iterating until every instance is separated and re-identified against the exemplars at its own scale. No fine-tuning, decoder, class labels, or full annotations are required. Because the exemplar defines the target \emph{scale and granularity}, a single prompt discovers \emph{every} instance of the prompted concept, both within an image and across images. As a proof of concept, we evaluate on two custom marine benchmarks --- coral and polyp segmentation --- and approach them training-free with roughly $40\times$ fewer parameters than a promptable foundation model. Polyps remain hard for all methods, but an oracle-foreground analysis shows that most of the gap is recoverable with a stronger foreground step, isolating it as the key bottleneck. We share our design, early results, and open questions to gather feedback.