--- dataset_info: features: - name: id dtype: string - name: image dtype: image - name: panoptic_mask 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': person '2': bicycle '3': car '4': motorcycle '5': airplane '6': bus '7': train '8': truck '9': boat '10': traffic light '11': fire hydrant '12': stop sign '13': parking meter '14': bench '15': bird '16': cat '17': dog '18': horse '19': sheep '20': cow '21': elephant '22': bear '23': zebra '24': giraffe '25': backpack '26': umbrella '27': handbag '28': tie '29': suitcase '30': frisbee '31': skis '32': snowboard '33': sports ball '34': kite '35': baseball bat '36': baseball glove '37': skateboard '38': surfboard '39': tennis racket '40': bottle '41': wine glass '42': cup '43': fork '44': knife '45': spoon '46': bowl '47': banana '48': apple '49': sandwich '50': orange '51': broccoli '52': carrot '53': hot dog '54': pizza '55': donut '56': cake '57': chair '58': couch '59': potted plant '60': bed '61': dining table '62': toilet '63': tv '64': laptop '65': mouse '66': remote '67': keyboard '68': cell phone '69': microwave '70': oven '71': toaster '72': sink '73': refrigerator '74': book '75': clock '76': vase '77': scissors '78': teddy bear '79': hair drier '80': toothbrush '81': banner '82': blanket '83': bridge '84': cardboard '85': counter '86': curtain '87': door-stuff '88': floor-wood '89': flower '90': fruit '91': gravel '92': house '93': light '94': mirror-stuff '95': net '96': pillow '97': platform '98': playingfield '99': railroad '100': river '101': road '102': roof '103': sand '104': sea '105': shelf '106': snow '107': stairs '108': tent '109': towel '110': wall-brick '111': wall-stone '112': wall-tile '113': wall-wood '114': water-other '115': window-blind '116': window-other '117': tree-merged '118': fence-merged '119': ceiling-merged '120': sky-other-merged '121': cabinet-merged '122': table-merged '123': floor-other-merged '124': pavement-merged '125': mountain-merged '126': grass-merged '127': dirt-merged '128': paper-merged '129': food-other-merged '130': building-other-merged '131': rock-merged '132': wall-other-merged '133': rug-merged - name: segments struct: - name: id list: int64 - name: area list: int64 - name: iscrowd list: bool - name: isthing list: bool - name: relations struct: - name: subject_index list: int64 - name: object_index list: int64 - name: predicate list: class_label: names: '0': __background__ '1': over '2': in front of '3': beside '4': 'on' '5': in '6': attached to '7': hanging from '8': on back of '9': falling off '10': going down '11': painted on '12': walking on '13': running on '14': crossing '15': standing on '16': lying on '17': sitting on '18': flying over '19': jumping over '20': jumping from '21': wearing '22': holding '23': carrying '24': looking at '25': guiding '26': kissing '27': eating '28': drinking '29': feeding '30': biting '31': catching '32': picking '33': playing with '34': chasing '35': climbing '36': cleaning '37': playing '38': touching '39': pushing '40': pulling '41': opening '42': cooking '43': talking to '44': throwing '45': slicing '46': driving '47': riding '48': parked on '49': driving on '50': about to hit '51': kicking '52': swinging '53': entering '54': exiting '55': enclosing '56': leaning on splits: - name: train num_bytes: 8178055254 num_examples: 46563 - name: validation num_bytes: 380822615 num_examples: 2186 download_size: 8560720927 dataset_size: 8558877869 configs: - config_name: default data_files: - split: train path: data/train-* - split: validation path: data/validation-* --- # Panoptic Scene Graph PSG（Panoptic Scene Graph）是面向全景场景图生成的数据集。每个场景图节点不仅包含对象类别和边界框，还通过 panoptic segmentation mask 与像素级区域对应。数据集同时覆盖 thing 和 stuff 类别。 本仓库将 RGB 图片和 panoptic PNG 直接嵌入 Parquet。通过 Hugging Face `Image` feature 加载后，两者都会解码为 Pillow 图像，不需要额外下载图片目录。 ## 数据集规模 | Split | 样本 | 对象 / segments | 关系 | | --- | ---: | ---: | ---: | | `train` | 46,563 | 513,243 | 261,666 | | `validation` | 2,186 | 24,910 | 13,705 | | **总计** | **48,749** | **538,153** | **275,371** | `validation` 对应完整 PSG 发布版中的官方 test split。完整发布版带有 test ground truth，因此这里保留其对象、mask 和关系标注，并改名为 Hugging Face 常用的 `validation`。部分样本没有关系，其关系数组为空。 PSG 中有少量不同样本引用同一张 COCO 图片；`id` 是 PSG 样本 ID，不应使用图片文件名替代。 ## 加载 ```python from datasets import load_dataset dataset = load_dataset("wliafe/PSG") sample = dataset["train"][0] image = sample["image"] # PIL.Image.Image panoptic_mask = sample["panoptic_mask"] # PIL.Image.Image print(sample["id"], image.size, panoptic_mask.size) ``` ## 字段 | 字段 | 类型 | 说明 | | --- | --- | --- | | `id` | `string` | PSG 样本 ID | | `image` | `Image` | 嵌入式 RGB 图片；源文件主要是 JPEG，少量 `.jpg` 实际为 PNG | | `panoptic_mask` | `Image` | 嵌入式 RGB 编码 panoptic PNG | | `width` / `height` | `int32` | 解码后 RGB 图片的实际像素尺寸 | | `boxes` | `List[[float32; 4]]` | 与 segments 平行的 `[x1, y1, x2, y2]` 边界框 | | `labels` | `List[ClassLabel]` | 与 segments 平行的对象类别 | | `segments.id` | `List[int64]` | panoptic mask 中的 segment ID | | `segments.area` | `List[int64]` | segment 的标注面积 | | `segments.iscrowd` | `List[bool]` | COCO crowd 标记 | | `segments.isthing` | `List[bool]` | thing/stuff 标记 | | `relations.subject_index` | `List[int64]` | 主语在当前 segment 数组中的零基索引 | | `relations.object_index` | `List[int64]` | 宾语在当前 segment 数组中的零基索引 | | `relations.predicate` | `List[ClassLabel]` | 有向关系的谓词类别 | `boxes`、`labels` 和四个 `segments` 数组长度相同。同一位置表示同一个场景图节点。三个 `relations` 数组长度也相同，同一位置共同表示一条有向关系。 ## Panoptic mask panoptic PNG 使用 RGB 三通道编码整数 segment ID： ```python import numpy as np rgb = np.asarray(sample["panoptic_mask"].convert("RGB"), dtype=np.int64) segment_id_map = ( rgb[..., 0] + 256 * rgb[..., 1] + 256 * 256 * rgb[..., 2] ) first_segment_id = sample["segments"]["id"][0] first_segment_mask = segment_id_map == first_segment_id print(first_segment_mask.sum()) ``` mask 中的非零 segment ID 与 `segments.id` 对应。背景像素的 ID 为 0，这与 taxonomy 中保留的 `__background__` 类别编号是两个不同概念。 ## 类别与关系 对象和谓词 taxonomy 都在索引 `0` 保留 `__background__`： - 对象前景类别编号为 `1`–`133`，依次对应 80 个 thing 类和 53 个 stuff 类。 - 谓词前景类别编号为 `1`–`56`。 - `subject_index` 和 `object_index` 是样本内部 segment 数组的位置，不是类别 ID。 ```python features = dataset["train"].features object_names = features["labels"].feature.names predicate_names = features["relations"]["predicate"].feature.names 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_]], ) ``` ## 坐标约定 `boxes` 使用实际图片像素坐标下的 `[x1, y1, x2, y2]` 格式，并位于图片边界内。`width` 和 `height` 与解码后的 RGB 图片及 panoptic mask 尺寸一致。 ## 使用限制 - 类别和谓词呈长尾分布，评估结果可能受高频类别主导。 - 场景图和分割标注可能包含遗漏、歧义或类别噪声。 - 本仓库不重新授予原始 COCO 图片版权；使用者应遵守 COCO 与 PSG 的许可和使用要求。 - 比较模型结果时，应确认采用相同 split、taxonomy 和 background 编号。 ## 引用 ```bibtex @inproceedings{yang2022panoptic, title={Panoptic Scene Graph Generation}, author={Yang, Jingkang and Ang, Yi Zhe and Guo, Zujin and Zhou, Kaiyang and Zhang, Wayne and Liu, Ziwei}, booktitle={European Conference on Computer Vision}, year={2022} } ```