Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
5
42
77227
5073960523_0 on 5073960523_1
5073960523_0 on 5073960523_2
5073960523_0 riding 5073960523_1
5073960523_0 has 5073960523_9
5073960523_0 has 5073960523_10
5073960523_0 wearing 5073960523_11
5073960523_0 right of 5073960523_12
5073960523_0 above 5073960523_12
5073960523_1 on 5073960523_2
5073960523_2 under 5073960523_0
5073960523_3 has 5073960523_4
5073960523_4 behind 5073960523_0
5073960523_6 in front of 5073960523_5
5073960523_6 behind 5073960523_0
5073960523_7 has 5073960523_8
5073960523_10 on 5073960523_1
5073960523_12 left of 5073960523_0
5073960523_12 has 5073960523_13
5073960523_12 has 5073960523_14
2752533795_0 left of 2752533795_1
2752533795_0 in 2752533795_17
2752533795_0 wearing 2752533795_18
2752533795_0 wearing 2752533795_19
2752533795_0 has 2752533795_14
2752533795_0 has 2752533795_20
2752533795_0 wearing 2752533795_21
2752533795_0 wearing 2752533795_22
2752533795_0 using 2752533795_1
2752533795_0 under 2752533795_2
2752533795_1 in front of 2752533795_0
2752533795_1 under 2752533795_2
2752533795_2 in front of 2752533795_3
2752533795_4 above 2752533795_5
2752533795_4 near 2752533795_10
2752533795_4 left of 2752533795_10
2752533795_6 in front of 2752533795_5
2752533795_7 left of 2752533795_8
2752533795_9 in front of 2752533795_3
2752533795_10 right of 2752533795_4
2752533795_10 near 2752533795_4
2752533795_11 left of 2752533795_12
2752533795_22 on 2752533795_0
2308524757_0 has 2308524757_1
2308524757_0 has 2308524757_2
2308524757_0 has 2308524757_5
2308524757_0 wearing 2308524757_9
2308524757_0 wearing 2308524757_10
2308524757_0 in front of 2308524757_8
2308524757_2 has 2308524757_3
2308524757_2 has 2308524757_4
2308524757_6 has 2308524757_7
2308524757_6 wearing 2308524757_11
2308524757_6 wearing 2308524757_12
2308524757_6 has 2308524757_17
2308524757_6 in front of 2308524757_8
2308524757_7 under 2308524757_12
2308524757_10 under 2308524757_5
2308524757_13 wearing 2308524757_14
2308524757_13 wearing 2308524757_16
2308524757_15 on 2308524757_14
2308524757_17 behind 2308524757_7
8481264570_0 on 8481264570_1
8481264570_0 near 8481264570_6
8481264570_1 near 8481264570_6
8481264570_1 on 8481264570_10
8481264570_2 near 8481264570_3
8481264570_2 wearing 8481264570_4
8481264570_2 on 8481264570_1
8481264570_2 using 8481264570_8
8481264570_2 near 8481264570_6
8481264570_3 has 8481264570_5
8481264570_3 on 8481264570_1
8481264570_3 near 8481264570_6
8481264570_6 near 8481264570_2
8481264570_6 near 8481264570_3
8481264570_7 above 8481264570_6
8481264570_7 at 8481264570_6
8481264570_7 over 8481264570_6
8481264570_9 on 8481264570_6
8481264570_10 at 8481264570_6
8481264570_10 has 8481264570_1
8481264570_10 near 8481264570_6
8481264570_10 near 8481264570_2
8481264570_11 near 8481264570_10
8519579814_0 in 8519579814_1
8519579814_0 near 8519579814_3
8519579814_1 has 8519579814_0
8519579814_2 in 8519579814_1
8519579814_2 on 8519579814_1
8519579814_3 wearing 8519579814_11
8519579814_3 eating 8519579814_6
8519579814_3 near 8519579814_5
8519579814_3 near 8519579814_0
8519579814_3 has 8519579814_8
8519579814_3 wearing 8519579814_17
8519579814_4 on 8519579814_5
8519579814_6 in 8519579814_6
8519579814_9 on 8519579814_10
8519579814_12 has 8519579814_13
End of preview. Expand in Data Studio

Spatial-VRD (Geometry-MMKG, data v0)

Inductive multimodal KG completion over scene graphs: every entity is one object box of a VRD image, 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?). Transductive control with the same triples: callmezuy/spatial-vrd-trans. Code: github.com/callmezuy/Geometry-MMKG (scripts/build_vrd_relation_review.py, scripts/clean_splits.py, scripts/extract_features.py).

Split Triples Entities
train 77,227 59,950
valid 8,614 6,694
test 21,563 16,908

93,809 entities (object boxes of 4,999 images), 52 relations. Entity names are <photo_id>_<k> (box k of the image).

Sources and cleaning

  • Annotations: the free-form sg_dataset annotations of VRD (Lu et al., ECCV 2016): sg_train_annotations.json, sg_test_annotations.json (1,310 raw predicates, object names and attributes). Images are not redistributed.
  • Relations: raw predicates mapped to the curated predicate vocabulary of callmezuy/spatial-vg plus a few VRD-only labels (left of, right of, touching, ...); 1.0% of triples dropped, some inverted. Mapping: relation_review.tsv.
  • Relations with < 10 triples dropped; duplicates within a split removed; no valid/test triple occurs in train.
  • Test = original VRD test images; valid = 10% of the train images (grouped by image).
  • relation_category.tsv: GEOMETRIC / POSSESSIVE / SEMANTIC / MISC per relation (our classification following the definitions of Zellers et al., CVPR 2018).
  • clean_stats.json: before / after counts of every step.

Features (*.pt, float32, row i = entity id i of entity2id.txt)

File Shape Content
ent_rgb_emb.pt [93809, 768] DINOv2-base pooler output of the box crop
ent_depth_emb.pt [93809, 768] ZipDepth (base, 384 px, per-image min-max) depth map of the whole image, box region cropped (roi_align) -> DINOv2-base pooler
ent_text_emb.pt [93809, 768] RoBERTa-base masked mean of entity2text.txt (attributes + object name)
ent_rgb_depth_mean_emb.pt [93809, 768] (rgb + depth) / 2
ent_rgb_depth_concat_emb.pt [93809, 1536] [rgb, depth]
ent_img_mask.pt [93809] bool True = entity has a usable box
ent_fill.pt dict vectors written into the rows of entities without a usable box

2 boxes (ids 61233, 61284) lie entirely outside their image (annotation error); they are treated as missing and filled with the mean rgb / depth feature of the train entities. Note that the text names the object class, so text alone predicts the class almost perfectly (feature_quality.json).

File formats

  • train/valid/test.txt: first line = count, then head \t relation \t tail; *2id.txt: OpenKE order head tail relation.
  • entity2id.txt, relation2id.txt: first line = count, then name \t id.
  • entity2text.txt: name \t text, no count line (23 extra lines name boxes that are not in entity2id.txt; ignore them).

Versions

v0 (18/11/2026) is the first version of this repo (earlier repos with this name were deleted). Pin it with snapshot_download("callmezuy/spatial-vrd", repo_type="dataset", revision="v0").

License

Derived from the VRD dataset (Lu et al., 2016); use under its terms.

Downloads last month
111