why the nb_classes = 90 about the COCO dataset in the datasets.py(184行)?

[znshi]

def build_dataset(is_train, data_set, args, gen_attn=False): transform = build_transform(is_train, args, gen_attn) dataset = None nb_classes = None

if data_set == 'VOC12':
    dataset = VOC12Dataset(data_root=args.data_path,
                           train=is_train, gen_attn=gen_attn, transform=transform)
    nb_classes = 20
elif data_set == 'VOC12MS':
    dataset = VOC12DatasetMS(img_name_list_path=args.img_ms_list, data_root=args.data_path,
                             scales=tuple(args.scales),
                             train=is_train, gen_attn=gen_attn, transform=transform)
    nb_classes = 20
elif data_set == 'COCO':
    dataset = COCOClsDataset(data_root=args.data_path,
                             train=is_train, gen_attn=gen_attn, transform=transform)
    nb_classes = 90
elif data_set == 'COCOMS':
    dataset = COCOClsDatasetMS(img_name_list_path=args.img_ms_list, data_root=args.data_path,
                               scales=tuple(args.scales),
                               train=is_train, gen_attn=gen_attn, transform=transform)
    nb_classes = 90

return dataset, nb_classes

[Yingyue-L]

Thank you for the question! Although the COCODataset actually has 80 classes, the official COCODataset annotations include 90 classes, of which 10 classes have never been observed. In our experiment, we directly used the official annotations without converting them to 80 classes. Therefore, the number of classes was set as 90.