그레프

  • 설명 :

Google RefExp 데이터 세트는 공개적으로 사용 가능한 MS-COCO 데이터 세트를 기반으로 하는 이미지의 개체에 대한 텍스트 설명 모음입니다. MS-COCO의 이미지 캡션은 전체 이미지에 적용되는 반면, 이 데이터 세트는 이미지 내의 단일 개체 또는 영역을 고유하게 식별할 수 있는 텍스트 설명에 중점을 둡니다. 이 백서에서 자세한 내용을 참조하십시오. 모호하지 않은 개체 설명의 생성 및 이해.

coco_train2014 폴더에는 모든 COCO 2014 훈련 이미지가 포함되어 있습니다.

  • 자동 캐시 ( 문서 ): 아니요

  • 분할 :

나뉘다
'train' 24,698
'validation' 4,650
  • 기능 구조 :
FeaturesDict({
    'image': Image(shape=(None, None, 3), dtype=uint8),
    'image/id': int64,
    'objects': Sequence({
        'area': int64,
        'bbox': BBoxFeature(shape=(4,), dtype=float32),
        'id': int64,
        'label': int64,
        'label_name': ClassLabel(shape=(), dtype=int64, num_classes=80),
        'refexp': Sequence({
            'raw': Text(shape=(), dtype=string),
            'referent': Text(shape=(), dtype=string),
            'refexp_id': int64,
            'tokens': Sequence(Text(shape=(), dtype=string)),
        }),
    }),
})
  • 기능 문서 :
특징 수업 모양 D타입 설명
풍모Dict
영상 영상 (없음, 없음, 3) uint8
이미지/ID 텐서 int64
사물 순서
개체/영역 텐서 int64
개체/bbox BBox기능 (4,) float32
객체/ID 텐서 int64
개체/레이블 텐서 int64
객체/label_name 클래스 레이블 int64
객체/참조 순서
객체/refexp/raw 텍스트
개체/refexp/referent 텍스트
개체/refexp/refexp_id 텐서 int64
객체/refexp/토큰 시퀀스(텍스트) (없음,)

심상

  • 인용 :
@inproceedings{mao2016generation,
  title={Generation and Comprehension of Unambiguous Object Descriptions},
  author={Mao, Junhua and Huang, Jonathan and Toshev, Alexander and Camburu, Oana and Yuille, Alan and Murphy, Kevin},
  booktitle={CVPR},
  year={2016}
}