Task-oriented RT-DETR adaptation for minute cigarette detection in complex laboratory scenes

Abstract Detecting minute cigarettes in complex laboratory surveillance scenes is challenging because of their small scale, partial occlusion, weak visual cues, cluttered backgrounds, and sensitivity to bounding-box deviations. This study develops a task-oriented adaptation of the Real-Time Detection Transformer (RT-DETR). Receptive Field Attention Convolution (RFAConv) is integrated into the ResNet-18 backbone, High-Low Frequency Attention (HiLo) is applied to high-level intra-scale interaction, and the existing Inner-MPDIoU formulation replaces GIoU as the intersection-over-union-based regression objective. Across three training runs, the complete model achieves an mAP@50 of $$89.6 \pm 0.4$$ % and an mAP@50:95 of $$50.2 \pm 0.2$$ %, compared with $$82.9 \pm 0.4$$ % and $$45.1 \pm 0.5$$ % for RT-DETR-R18. On an independent 150-image hold-out set, it achieves $$85.8 \pm 0.2$$ % mAP@50 and $$46.2 \pm 0.1$$ % mAP@50:95, exceeding the baseline values of $$77.2 \pm 0.2$$ % and $$41.7 \pm 0.3$$ %. In batch-one end-to-end inference on an NVIDIA GeForce RTX 4060 Ti, the model achieves 18 ms per frame and 56 FPS, with 3.4 GB peak GPU-memory occupancy. Although computation increases from 58.3 to 98.3 GFLOPs, the results demonstrate improved detection and hold-out performance while retaining practical workstation-level inference efficiency.