Leveraging Foundation Models to Improve Insect Localization in Field Images

The automatic detection of insects in natural environments is a critical challenge for ecological monitoring. While detection on homogeneous backgrounds is largely solved, localizing insects in "wild" settings with complex textures remains an open task. To address this limitation, this thesis proposes a novel zero-shot detection pipeline leveraging Grounding DINO and Segment Anything Model 3 (SAM3). After evaluating multiple prompting strategies, a lightweight supervised binary discriminator is introduced as a post-processing step to re-score predictions based on the presence of insects. Benchmarked across diverse full-frame and close-up datasets, the proposed pipeline demonstrates exceptional out-of-distribution capabilities, consistently outperforming state of-the-art supervised models in single flower monitoring scenarios. Finally, the limitations of this approach are discussed, particularly regarding the computational cost of foundation models inference.