metadata
dataset_info:
features:
- name: id
dtype: string
- name: image
dtype: image
- name: width
dtype: int32
- name: height
dtype: int32
- name: boxes
list:
list: float32
length: 4
- name: labels
list:
class_label:
names:
'0': __background__
'1': airplane
'2': animal
'3': arm
'4': bag
'5': banana
'6': basket
'7': beach
'8': bear
'9': bed
'10': bench
'11': bike
'12': bird
'13': board
'14': boat
'15': book
'16': boot
'17': bottle
'18': bowl
'19': box
'20': boy
'21': branch
'22': building
'23': bus
'24': cabinet
'25': cap
'26': car
'27': cat
'28': chair
'29': child
'30': clock
'31': coat
'32': counter
'33': cow
'34': cup
'35': curtain
'36': desk
'37': dog
'38': door
'39': drawer
'40': ear
'41': elephant
'42': engine
'43': eye
'44': face
'45': fence
'46': finger
'47': flag
'48': flower
'49': food
'50': fork
'51': fruit
'52': giraffe
'53': girl
'54': glass
'55': glove
'56': guy
'57': hair
'58': hand
'59': handle
'60': hat
'61': head
'62': helmet
'63': hill
'64': horse
'65': house
'66': jacket
'67': jean
'68': kid
'69': kite
'70': lady
'71': lamp
'72': laptop
'73': leaf
'74': leg
'75': letter
'76': light
'77': logo
'78': man
'79': men
'80': motorcycle
'81': mountain
'82': mouth
'83': neck
'84': nose
'85': number
'86': orange
'87': pant
'88': paper
'89': paw
'90': people
'91': person
'92': phone
'93': pillow
'94': pizza
'95': plane
'96': plant
'97': plate
'98': player
'99': pole
'100': post
'101': pot
'102': racket
'103': railing
'104': rock
'105': roof
'106': room
'107': screen
'108': seat
'109': sheep
'110': shelf
'111': shirt
'112': shoe
'113': short
'114': sidewalk
'115': sign
'116': sink
'117': skateboard
'118': ski
'119': skier
'120': sneaker
'121': snow
'122': sock
'123': stand
'124': street
'125': surfboard
'126': table
'127': tail
'128': tie
'129': tile
'130': tire
'131': toilet
'132': towel
'133': tower
'134': track
'135': train
'136': tree
'137': truck
'138': trunk
'139': umbrella
'140': vase
'141': vegetable
'142': vehicle
'143': wave
'144': wheel
'145': window
'146': windshield
'147': wing
'148': wire
'149': woman
'150': zebra
- name: attributes
list:
list:
class_label:
names:
'0': __background__
'1': white
'2': black
'3': blue
'4': green
'5': red
'6': brown
'7': yellow
'8': small
'9': large
'10': wooden
'11': silver
'12': orange
'13': grey
'14': tall
'15': long
'16': dark
'17': pink
'18': standing
'19': round
'20': tan
'21': glass
'22': here
'23': wood
'24': open
'25': purple
'26': short
'27': plastic
'28': parked
'29': sitting
'30': walking
'31': striped
'32': brick
'33': young
'34': gold
'35': old
'36': hanging
'37': empty
'38': 'on'
'39': bright
'40': concrete
'41': cloudy
'42': colorful
'43': one
'44': beige
'45': bare
'46': wet
'47': light
'48': square
'49': closed
'50': stone
'51': shiny
'52': thin
'53': dirty
'54': flying
'55': smiling
'56': painted
'57': thick
'58': part
'59': sliced
'60': playing
'61': tennis
'62': calm
'63': leather
'64': distant
'65': rectangular
'66': looking
'67': grassy
'68': dry
'69': cement
'70': leafy
'71': wearing
'72': tiled
'73': man's
'74': baseball
'75': cooked
'76': pictured
'77': curved
'78': decorative
'79': dead
'80': eating
'81': paper
'82': paved
'83': fluffy
'84': lit
'85': back
'86': framed
'87': plaid
'88': dirt
'89': watching
'90': colored
'91': stuffed
'92': clean
'93': in the picture
'94': steel
'95': stacked
'96': covered
'97': full
'98': three
'99': street
'100': flat
'101': baby
'102': black and white
'103': beautiful
'104': ceramic
'105': present
'106': grazing
'107': sandy
'108': golden
'109': blurry
'110': side
'111': chocolate
'112': wide
'113': growing
'114': chrome
'115': cut
'116': bent
'117': train
'118': holding
'119': water
'120': up
'121': arched
'122': metallic
'123': spotted
'124': folded
'125': electrical
'126': pointy
'127': running
'128': leafless
'129': electric
'130': in background
'131': rusty
'132': furry
'133': traffic
'134': ripe
'135': behind
'136': laying
'137': rocky
'138': tiny
'139': down
'140': fresh
'141': floral
'142': stainless steel
'143': high
'144': surfing
'145': close
'146': 'off'
'147': leaning
'148': moving
'149': multicolored
'150': woman's
'151': pair
'152': huge
'153': some
'154': background
'155': chain link
'156': checkered
'157': top
'158': tree
'159': broken
'160': maroon
'161': iron
'162': worn
'163': patterned
'164': ski
'165': overcast
'166': waiting
'167': rubber
'168': riding
'169': skinny
'170': grass
'171': porcelain
'172': adult
'173': wire
'174': cloudless
'175': curly
'176': cardboard
'177': jumping
'178': tile
'179': pointed
'180': blond
'181': cream
'182': four
'183': male
'184': smooth
'185': hazy
'186': computer
'187': older
'188': pine
'189': raised
'190': many
'191': bald
'192': snow covered
'193': skateboarding
'194': narrow
'195': reflective
'196': rear
'197': khaki
'198': extended
'199': roman
'200': american
- name: relations
struct:
- name: subject_index
list: int64
- name: object_index
list: int64
- name: predicate
list:
class_label:
names:
'0': __background__
'1': above
'2': across
'3': against
'4': along
'5': and
'6': at
'7': attached to
'8': behind
'9': belonging to
'10': between
'11': carrying
'12': covered in
'13': covering
'14': eating
'15': flying in
'16': for
'17': from
'18': growing on
'19': hanging from
'20': has
'21': holding
'22': in
'23': in front of
'24': laying on
'25': looking at
'26': lying on
'27': made of
'28': mounted on
'29': near
'30': of
'31': 'on'
'32': on back of
'33': over
'34': painted on
'35': parked on
'36': part of
'37': playing
'38': riding
'39': says
'40': sitting on
'41': standing on
'42': to
'43': under
'44': using
'45': walking in
'46': walking on
'47': watching
'48': wearing
'49': wears
'50': with
splits:
- name: train
num_bytes: 8372499067
num_examples: 57723
- name: validation
num_bytes: 782338184
num_examples: 5000
- name: test
num_bytes: 3501331914
num_examples: 26446
download_size: 12653741212
dataset_size: 12656169165
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
VG150
VG150 是从 Visual Genome 派生的场景图生成 （Scene Graph Generation, SGG）基准数据集，包含 150 个对象类别、50 个 关系谓词类别和 200 个属性类别。图片直接嵌入 Parquet。
该数据集适用于场景图生成、场景图检测、视觉关系检测和相关多模态研究。
数据集规模
| Split | 图片数量 | 用途 |
|---|---|---|
train |
57,723 | 模型训练 |
validation |
5,000 | 验证与模型选择 |
test |
26,446 | 最终评估 |
| 总计 | 89,169 | 仅包含同时具有对象和关系标注的图片 |
所有 split 均包含完整的对象、属性和关系标注。
加载
from datasets import load_dataset
dataset = load_dataset("wliafe/VG150")
print(dataset)
sample = dataset["train"][0]
image = sample["image"] # PIL.Image.Image
image.show()
图片字节已经嵌入 Parquet，因此不需要额外下载或拼接 Visual Genome 图片目录。
image 使用 Hugging Face Image feature，默认读取时会解码为 Pillow 图像。
如需取得类别名称，可以读取 ClassLabel feature：
features = dataset["train"].features
object_names = features["labels"].feature.names
predicate_names = features["relations"]["predicate"].feature.names
attribute_names = features["attributes"].feature.feature.names
sample = dataset["train"][0]
object_labels = [object_names[index] for index in sample["labels"]]
predicate_labels = [
predicate_names[index]
for index in sample["relations"]["predicate"]
]
attribute_labels = [
[attribute_names[index] for index in object_attributes]
for object_attributes in sample["attributes"]
]
数据字段
| 字段 | 类型 | 说明 |
|---|---|---|
id |
string |
Visual Genome 图片 ID |
image |
Image |
嵌入 Parquet 并可直接解码的图片 |
width |
int32 |
实际 JPEG 宽度，单位为像素 |
height |
int32 |
实际 JPEG 高度，单位为像素 |
boxes |
List[[float32; 4]] |
与对象平行的 [x1, y1, x2, y2] 边界框 |
labels |
List[ClassLabel] |
与 boxes 平行的对象类别 |
attributes |
List[List[ClassLabel]] |
每个对象对应的零个或多个属性 |
relations.subject_index |
List[int64] |
每条关系的主语对象索引 |
relations.object_index |
List[int64] |
每条关系的宾语对象索引 |
relations.predicate |
List[ClassLabel] |
每条关系的谓词类别 |
boxes、labels 和 attributes 的外层长度相同。三个关系数组也具有相同
长度；相同位置的主语索引、宾语索引和谓词共同构成一条有向关系：
sample = dataset["train"][0]
for subject, object_, predicate in zip(
sample["relations"]["subject_index"],
sample["relations"]["object_index"],
sample["relations"]["predicate"],
):
print(
object_names[sample["labels"][subject]],
predicate_names[predicate],
object_names[sample["labels"][object_]],
)
类别编号
对象、谓词和属性 taxonomy 均在索引 0 保留 __background__：
- 对象前景类别编号为
1–150。 - 谓词前景类别编号为
1–50。 - 属性前景类别编号为
1–200。
没有属性的对象对应空列表，而不是包含 __background__ 的列表。
边界框与图片尺寸
boxes 使用实际图片像素坐标下的 [x1, y1, x2, y2] 格式，坐标位于图片
边界内。width 和 height 与 image 的实际尺寸一致。关系端点使用当前样本
对象数组中的零基索引。
本仓库提供可直接加载的 Parquet 数据，不包含原始 H5 或 JSON 文件。对象、关系和 属性标注可能继承 Visual Genome 中的类别偏差、长尾分布、歧义和错误。
使用限制
- VG150 只保留高频类别，不能代表 Visual Genome 的完整开放词汇分布。
- 关系类别分布高度不均衡，模型结果可能被少数高频谓词主导。
boxes是对象边界框，不是实例分割轮廓。- 本仓库不重新授予原始图片版权。使用者应同时遵守 Visual Genome 的使用条款， 并确认原始图片适用于自己的研究或发布场景。
- 比较论文结果时，应确认对方使用相同的 VG150 taxonomy、过滤规则和 split。
引用
使用本数据集时，请引用 Visual Genome：
@article{krishna2017visual,
title={Visual Genome: Connecting Language and Vision Using Crowdsourced Dense Image Annotations},
author={Krishna, Ranjay and Zhu, Yuke and Groth, Oliver and Johnson, Justin and
Hata, Kenji and Kravitz, Joshua and Chen, Stephanie and
Kalantidis, Yannis and Li, Li-Jia and Shamma, David A. and
Bernstein, Michael S. and Fei-Fei, Li},
journal={International Journal of Computer Vision},
volume={123},
number={1},
pages={32--73},
year={2017}
}
标准 VG150 场景图划分也常用于 Neural Motifs：
@inproceedings{zellers2018neural,
title={Neural Motifs: Scene Graph Parsing with Global Context},
author={Zellers, Rowan and Yatskar, Mark and Thomson, Sam and Choi, Yejin},
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
year={2018}
}