Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
5
26
319069
2 wearing 11
5 on 6
2 has 7
9 has 10
12 has 11
2 wearing 13
12 has 14
12 has 15
4 parked on 1
9 parked on 8
4 on 16
2 wearing 7
4 near 18
12 wearing 11
7 on 2
2 in 7
18 near 8
4 behind 2
18 near 1
12 wearing 14
6 with 22
24 parked on 25
26 near 25
27 near 25
28 near 25
36 parked on 25
37 near 25
26 along 25
27 along 25
28 along 25
38 near 24
43 near 42
41 on 42
44 on 42
45 has 47
45 with 47
50 on 46
53 on 52
45 sitting at 55
57 on 60
61 against 57
56 near 63
66 behind 60
61 on 60
66 on 58
72 on 77
73 near 79
72 on 70
77 on 70
73 against 70
80 against 70
84 on 85
85 has 84
86 on 87
88 near 83
88 against 83
92 on 83
91 on 83
89 on 83
93 in 94
94 on 83
95 on 83
90 on 83
92 near 89
97 under 83
85 has 98
94 of 93
99 on 100
102 with 103
104 has 105
102 wearing 106
107 near 104
104 on 111
112 above 100
102 sitting at 101
102 sitting on 114
116 wearing 121
123 on 124
125 on 126
127 on 128
126 on 117
116 has 120
131 on 128
116 wearing 133
116 has 121
135 of 131
136 has 137
140 on 141
137 in 136
154 on 152
147 against 143
154 under 155
148 on 147
150 on 153
150 with 158
159 on 160
163 on 162
164 has 165
167 on 166
End of preview. Expand in Data Studio

Spatial-VG (Geometry-MMKG, data v0)

Inductive multimodal KG completion over Visual Genome scene graphs: every entity is one object box, triples are the annotated relations between boxes, and the split is by image, so valid / test entities never occur in train. Built for the Geometry-MMKG thesis (does monocular depth help multimodal KG completion?). Code: github.com/callmezuy/Geometry-MMKG (scripts/clean_splits.py, scripts/extract_features_v4.py, notebooks/extract-vg-features.py).

Split Triples Entities in triples
train 319,069 389,413
valid 26,372 33,049
test 137,984 176,343

1,441,984 entities in entity2id.txt (every box of VG-SGG.h5; entity id = box index, so most boxes are in no triple), 50 relations, 108,073 images.

Sources and cleaning

  • VG-SGG.h5, VG-SGG-dicts.json: VG150-curated, variant no_part_whole (Neau et al., ICCVW 2023, Maelic/VG_curated), not the standard VG150 of Xu et al. (2017). Images: Visual Genome 1.4 (Krishna et al., 2017); not redistributed here.
  • Split = the official image split of the H5 file; duplicates within a split removed (train 103,092, valid 8,723, test 40,935); no valid/test triple occurs in train. clean_stats.json has every count.
  • relation_category.tsv: GEOMETRIC / POSSESSIVE / SEMANTIC / MISC per relation (our classification following the definitions of Zellers et al., CVPR 2018).

Features

Not in this commit yet. The v0 features (ent_rgb_emb.pt, ent_depth_emb.pt, ent_text_emb.pt, ent_img_mask.pt, ent_fill.pt, feature_stats.json, feature_quality.json) are being extracted with notebooks/extract-vg-features.py and will be added before the tag v0 is created. Same recipe as callmezuy/spatial-vrd: DINOv2-base on the box crop (rgb), ZipDepth depth map of the whole image cropped to the box and encoded by DINOv2-base (depth), RoBERTa-base on the class name (text). This repo has no mean / concat tables.

File formats

  • train/valid/test.txt: first line = count, then head \t relation \t tail (entity names are box indices); *2id.txt: OpenKE order head tail relation.
  • entity2id.txt, relation2id.txt: first line = count, then name \t id.
  • entity2text.txt: first line = count, then name \t class name.

Versions

v0 is the first version of this repo (earlier repos with this name were deleted); the tag is created once the features are in. Pin it with snapshot_download("callmezuy/spatial-vg", repo_type="dataset", revision="v0").

License

Derived from Visual Genome (CC BY 4.0) and VG150-curated; use under the terms of those sources.

Downloads last month
109