Small-Pollinator Detection in Cluttered Field Video

2026-07-24

AI Summary

Detecting small, visually similar pollinator insects in complex field videos—often occluded or blurred—is highly challenging. This study systematically evaluates the performance of models such as YOLO and RF-DETR under single-GPU constraints across varying input resolutions, while exploring strategies including sliced inference, class-gated fusion, size-aware ensemble routing, and temporal post-processing. Experimental results demonstrate that RF-DETR Large with high-resolution inputs (up to 1344 pixels) significantly improves detection of small and rare species, such as bumblebees and moths, achieving an mAP50:95 of 0.405—the best single-model result on the BuzzSpot challenge to date. The work further identifies feature-level temporal aggregation as a critical direction for enhancing detection robustness in future research.

Abstract

Detecting pollinators in field video is challenging: targets are small, visually similar, and observed against cluttered vegetation under blur and occlusion. We present a systematic empirical study of small-pollinator detection under a practical single-GPU compute budget. Using the BuzzSpot challenge dataset, we compare YOLO and RF-DETR models across input resolutions and evaluate sliced inference, class-gated fusion, size-routed ensembling, and post-hoc temporal processing. RF-DETR Large at 1344-pixel resolution achieved our best hidden-test result, reaching 0.405 mAP50:95 and outperforming the 1120-pixel model (0.379) and the best single-model YOLO26m baseline (0.366). The strongest gains came from adopting RF-DETR and increasing its input resolution, indicating that detector choice and input resolution were more effective levers than added inference-time complexity; the resolution gain was strongest for small objects and the rarer bumblebee and moth classes. Sliced-inference fusion, size-routed ensembling, and warm-started 1536-pixel continuation did not surpass this result, while post-hoc temporal processing did not improve the leaked diagnostic evaluation. Error analysis identified bee-hoverfly discrimination as the clearest remaining bottleneck: neighboring frames rarely supplied correctly classified hoverfly evidence for post-hoc correction. These findings motivate learned feature-level temporal aggregation before the final classification decision.

Problem

• small-pollinator detection
• cluttered field video
• object detection
• occlusion
• visual similarity

Innovation

• small-object detection
• RF-DETR
• input resolution scaling
• pollinator monitoring
• temporal feature aggregation