SOCO

SOCO is a semanti object correspondence dataset with images, per-view keypoint annotations, pair annotations that for evaluation of visual features. Evaluation code is available at the OmniProbe repo.

Repository Layout

GenIntelLab/SOCO
  Images.zip                # -> Images/<category>/*.JPEG
  KeypointAnnotations.zip   # -> KeypointAnnotations/<category>/*.json
  PairAnnotations.zip       # -> PairAnnotations/{intra,cross,trainsplits/{train,test}}/<category>/*.json
  Metadata/
    filename_mapping.json
    keypoint_taxonomy.json
  README.md

After unzipping, the dataset tree is:

SOCOv1/
  Images/<category>/*.JPEG
  KeypointAnnotations/<category>/*.json
  PairAnnotations/
    intra/<category>/*.json
    cross/<category>/*.json
    trainsplits/
      train/<category>/*.json
      test/<category>/*.json
  Metadata/
    filename_mapping.json
    keypoint_taxonomy.json

Contents

• Images.zip: rendered object images, organized by category.
• KeypointAnnotations.zip: per-view keypoint annotations, organized by category.
• PairAnnotations.zip: image-pair files for intra-category (intra), cross-category (cross), and the predefined train/test splits (trainsplits).
• Metadata/: keypoint taxonomy and filename mapping (shipped unzipped).

This release contains 100 categories, 4,000 images (40 per category), 4,000 keypoint annotation files, and 60,000 pair annotation files (20,000 intra-category, 20,000 cross-category, and a 10,000 / 10,000 intra-category train / test split).

Download

Install the Hub client:

pip install -U huggingface_hub

Download the repository and unpack the archives:

huggingface-cli download GenIntelLab/SOCO --repo-type dataset --local-dir SOCOv1
cd SOCOv1 && for z in *.zip; do unzip -q "$z" && rm "$z"; done

Equivalently in Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="GenIntelLab/SOCO",
    repo_type="dataset",
    local_dir="SOCOv1",
)
# then unzip Images.zip, KeypointAnnotations.zip, PairAnnotations.zip inside SOCOv1/

Using the data

After extraction the folder layout maps directly onto a dataset-root path: point your loader at the SOCOv1/ directory and read images from Images/<category>/ and pairs from the relevant PairAnnotations/ subfolder. For evaluation, use PairAnnotations/intra (within-category pairs). For training and evaluating a probe, use the predefined PairAnnotations/trainsplits/train and PairAnnotations/trainsplits/test splits. Each pair file references two views and their corresponding keypoints. Per-view keypoints are also available under KeypointAnnotations/<category>/.

Citation

@misc{duenkel2026soco,
  title         = {SOCO: Benchmarking Semantic Object Correspondence in Vision Foundation Models},
  author        = {D{\"u}nkel, Olaf and Sunagad, Basavaraj and Wang, Haoran and
                   Hoffmann, David T. and Theobalt, Christian and Kortylewski, Adam},
  year          = {2026},
  eprint        = {2605.31597},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2605.31597}
}