PACE: A Large-Scale Dataset with Pose Annotations in Cluttered Environments (ECCV 2024)

PACE is a large-scale benchmark for object pose estimation and tracking in cluttered, real-world scenes, covering both instance-level and category-level tasks.

• 55K real frames with 258K pose annotations across 300 videos
• 238 objects from 43 categories, rigid and articulated
• Annotated with a calibrated 3-camera RGB-D capture rig
• PACESim: 100K photo-realistic simulated frames with 2.4M annotations across 931 objects

Links:

• Paper: arXiv:2312.15130 · Hugging Face paper page · ECCV 2024 PDF · Supplementary
• Code, evaluation and annotation tools: https://github.com/qq456cvb/PACE
• Project page: https://qq456cvb.github.io/projects/pace

Files

File	Contents
test_chunk_a[a-f]	Real-world test set, split into chunks
train_pbr_cat_chunk_a[a-g]	Simulated (PBR) training data, category level, split into chunks
train_pbr_inst_chunk_a[a-g]	Simulated (PBR) training data, instance level, split into chunks
val_inst.tar.gz	Validation set, instance level
val_pbr_cat.tar.gz	Simulated (PBR) validation set, category level
models.tar.gz	Scanned object meshes
models_eval.tar.gz	Uniformly sampled point clouds for evaluation (e.g. Chamfer distance)
models_nocs.tar.gz	Meshes with vertices colored by NOCS coordinates
model_splits/	Object IDs for the train/val/test splits, category and instance level
test_targets_bop19.json	BOP-style test targets
baseline_results/	Baseline predictions and category-level ground truth (catpose_gts_test.pkl)

The data follows the BOP format. Each scene folder holds rgb, depth, mask, mask_visib and (for category level) rgb_nocs images, plus scene_camera.json, scene_gt.json, scene_gt_info.json and COCO-format 2D boxes and masks. See the GitHub README for the full format description.

Download

The full dataset is about 870 GB. Download only the parts you need with --include:

pip install -U "huggingface_hub[cli]"

# Example: real test set and object models only (about 73 GB)
hf download qq456cvb/PACE --repo-type dataset --local-dir dataset/pace \
  --include "test_chunk_*" --include "models*.tar.gz" \
  --include "model_splits/*" --include "test_targets_bop19.json"

Merge the chunked archives, then extract everything under dataset/pace:

cd dataset/pace
cat test_chunk_* > test.tar.gz
cat train_pbr_cat_chunk_* > train_pbr_cat.tar.gz     # if downloaded
cat train_pbr_inst_chunk_* > train_pbr_inst.tar.gz   # if downloaded
for f in *.tar.gz; do tar -xzf "$f"; done

Benchmark evaluation

Evaluation code for instance-level pose estimation (BOP toolkit) and category-level pose estimation is in the GitHub repository. Baseline results are in baseline_results/: CosyPose, GDRNPP, PPF and SurfEmb for instance level, and ANCSH, CPPF++, HS-Pose, NOCS, SAR-Net and SGPA for category level.

License

MIT, except:

• Models with IDs 693–1260 come from Sketchfab under CC BY 4.0. The original post for each is at https://sketchfab.com/3d-models/${OBJ_IDENTIFIER}; the identifier is in models_info.json.
• Models 1165 and 1166 come from GrabCAD and are subject to the GrabCAD license.

Citation

@inproceedings{you2024pace,
  title={PACE: A Large-Scale Dataset with Pose Annotations in Cluttered Environments},
  author={You, Yang and Xiong, Kai and Yang, Zhening and Huang, Zhengxiang and Zhou, Junwei and Shi, Ruoxi and Fang, Zhou and Harley, Adam W. and Guibas, Leonidas and Lu, Cewu},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2024},
  organization={Springer}
}