Get started
Work

A Three-Stage Effective Receptive Field (ERF)-Guided Training Framework for Small-Object Detection in Autonomous Driving Systems

Year: 2026
Type: conference-paper
Abstract: Small object detection is a key bottleneck for autonomous driving perception systems, as lightweight detectors such as YOLO have a structurally limited effective field of view and an inherent tendency to reject low-resolution targets. Many recent approaches to compensate for this weakness rely on heavy modules-from emission-based reconstruction to multi-stage heads and transformer blocks-but their computational overhead prevents them from being used in real-time applications. This paper presents a small object-aware, Effective Receptive Field (ERF)-aware training scheme that significantly improves the model's sensitivity to small objects on the BDD100K dataset without any changes to the YOLOv8 architecture and without increasing the inference cost. The proposed strategy consists of three complementary components: (1) basic training at standard resolution to stabilize the effective behavior of the receptive field, (2) a higher resolution enhancement stage with increased spatial weights and enhanced composite data aiming to reallocate gradients to areas associated with small objects, and (3) gradual spine stabilization and fine-tuning of the head to correct classification errors without reducing overall features. Experiments on the BDD100K evaluation image show that this curriculum increases mAP50 from 0.583 to 0.604 and mAP50-95 from 0.340 to 0.353, while fully preserving the inference speed. Detailed analysis of the PR curves, F1-Confidence, and error matrices shows that this improvement is mainly due to the real increase in recall in small classes - including traffic lights, traffic signs, bicycles, and motorcycles - and the reduction of background collapse.
DOI: 10.1109/icee70628.2026.11710604
Language: English
Open Access status: gold
OpenAlex ID: W7220901588
FWCI:
Cites: 5
Cited by:
Related to: 10