DenseDetect: Exploring End-to-End Human Detection Models for Dense and Occluded Scenes

Unlike detection tasks in common scenarios, human detection in dense scenes faces challenges such as severe occlusion between individuals, leading to feature loss, and the difficulty of detecting small targets, which makes existing object detection models even powerful MLLMs unsuitable for this task. DenseDetect was proposed to alleviate these challenges integrating a multi-scale feature extraction network with two newly designed inference detection heads. DenseDetect outperforms classical object detection models on benchmarks such as CrowdHuman, surpassing metrics like , and . Additionally, even when fine-tuned on crowd detection datasets, our model retains its general object detection capabilities, showing no degradation and even slight improvements across metrics on the COCO dataset.