Semantic Person Segmentation

Binary classification is the most suitable task considering the common use cases in MCUs. Numerous datasets for image classification have been proposed. The Visual Wake Words (VWW) dataset, which is derived from the COCO dataset, distinguishes between ‘w/ person’ and ‘w/o person’ and is designed for object detection on MCUs. Therefore, datasets for binary classification and object detection exist. However, the dataset for binary classification has not been proposed for the semantic segmentation task. Segmentation requires significant memory because the need to generate a semantic mask makes deployment on MCUs challenging. However, HARD demonstrated that the semantic segmentation model can be deployed on MCUs. Therefore,we propose a new dataset named Semantic Person Segmentation (SPS), which segments “person pixels”. This dataset filters the labels from the publicly available PASCAL VOC dataset to provide a dataset of 2,301 training and validation images. The semantic mask labels will assign a value of 1 to pixel regions where “person” class objects are present and a value of 0 to pixel regions where “person” class objects are not present.