A reproducible benchmark for trophallaxis state detection in Camponotus fellah using YOLO26 and RF-DETR

Abstract Automated detection of trophallaxis in dense ex situ ant footage could enable large-scale quantitative analysis of social interactions, but remains challenging due to small object size, severe occlusion, and strong class imbalance. We present a reproducible benchmark for behavior-aware trophallaxis detection in Camponotus fellah under a unified COCO evaluation protocol. We compare a convolutional one-stage detector (YOLO26) and a transformer-based detector (RF-DETR) under comparable training and inference settings, reporting detection accuracy and computational performance. On a 2511-image dataset with track-majority splits, RF-DETR Small achieves higher detection quality (test mAP@[0.5:0.95] 0.467 vs. 0.364 for YOLO26n), with improved precision and recall at IoU ≥ 0.5, while YOLO26n provides approximately 2.1× higher throughput (≈ 49 vs. ≈ 18 FPS on NVIDIA RTX 4070). Per-class analysis indicates the largest gains on the minority trophallaxis class, consistent with improved interaction-state sensitivity; causal attribution to a specific architectural mechanism is not established in this benchmark study. Complementary experiments on an auxiliary dense-ant benchmark and temporal smoothing ablations indicate consistent accuracy–latency trade-offs and limited impact of post-processing on overall mAP. We additionally report a Faster R-CNN R50-FPN baseline, which achieves intermediate performance between YOLO26n and RF-DETR under the same evaluator. We publicly release the dataset, split manifests, evaluation scripts, and model checkpoints as a reproducible benchmark for this task. The study does not claim field-ready deployment or full behavioral understanding; it establishes frame-level detector baselines under controlled ex situ conditions. These results provide initial practical guidance: RF-DETR when interaction-state sensitivity is critical, and YOLO26 when real-time constraints dominate.