Code and benchmark
Source and access
Relevancy to Gratheon
The paper documents both successful and unsuccessful interventions for detecting very small pollinators in realistic video. This is valuable for Gratheon because it shows that increasing detector resolution and choosing a stronger architecture can matter more than adding slicing, routing, or post-processing complexity.
The repository includes fifteen experiments, hidden-test metrics, and integrity hashes, providing an unusually reproducible comparison. Its class-level analysis also exposes bee-hoverfly confusion as a concrete validation target for pollination monitoring, while the negative temporal result suggests that useful video context may need to be learned before classification rather than fused after independent frame detections.