Overlapping object detection with adaptive Gaussian sample division and asymmetric weighted loss

Existing deep learning based detectors are mostly designed for scenes with sparsely distributed objects. However, in certain scenarios such as dense crowds, objects often overlap severely. The dense anchor arrangement in anchor-based detectors is not quite suitable for the overlapping object detection. Anchor-free detectors have the potential to achieve high-performance in overlapping object detection, but troubled by the extreme imbalance of positive and negative samples. To this end, we propose an anchor-free overlapping object detector. Our adaptive Gaussian sample division (AGSD) can effectively allocate positive and negative samples with clear semantics to overlapping objects. Secondly, asymmetric weighted loss (AW Loss) adapts to continuous positive and negative sample values, thereby improving the classification ability of the detector. Lastly, our global location distribution head (GLD head) can introduce the supervision of overlapping object distributions. To verify the effectiveness of our method, we construct a large-scale high-quality overlapping object detection dataset containing 6,173 images and 17,725 annotations. Compared with mainstream object detector, our method achieves the best performance of AP50 at 96.71%.