Investigation of the Performance of Different Loss Function Types Within Deep Neural Anchor-Free Object Detectors

Ala'a Alshubbak, Daniel Görges

Abstract: In this paper, an investigation of different IoU loss functions and a spatial attention mechanism within anchor-free object detectors is presented. Two anchor-free dense predictor models are studied: FASF and FCOS models. The models are tested on two different datasets: the benchmark COCO dataset and a small dataset called OPEDD. The results show that some loss functions and using the attention mechanism outperform their original counterparts for both the huge multi-class COCO dataset and the small unity-class dataset of OPEDD. The proposed structure is tested over different backbones: ResNet-50, ResNet-101, and ResNeXt-101. The accuracy of basic models trained over the coco dataset improves by 1.3% and 1.6% mAP for the FSAF and FCOS models based on ResNet-50, respectively. On the other hand, it increases by 2.3% and 15.8% for the same models when trained on the OPEDD dataset. The effect is interpreted using a saliency map.