Instance Segmentation based Object Detection with Enhanced Path Aggregation Network
Ade Indra Onthoni, Prasan Kumar Sahoo · 2023
Multi-scale networks rely heavily on the aggregation process to aggregate feature information into other feature maps. However, it has disadvantages in aggregating information, where the localization information becomes inconsistent. An Enhancing Path Aggregation Network through concatenation and attention mechanism is proposed here to improve the aggregation path for generating the feature maps in multi-scale network. This approach helps to keep information consistently during the process of aggregation. The proposed model is evaluated using the COCO dataset, which achieves significant improvements on several metrics.