PLAN-Lab/MixedParts

Mixed Parts

Mixed Parts is the annotation bundle used by CALICO for part-focused semantic co-segmentation. It contains multi-image object-part comparison samples curated from ADE20KPart234, PartImageNet, and PACO-LVIS image assets.

This repository contains annotations only. Users must download the original image datasets separately from their upstream sources and overlay these annotations into the expected CALICO directory structure.

Files

mixed_parts_train.json
mixed_parts_val.json
mixed_parts_test.json
ADE20KPart234/ade20k_instance_train_mixed_parts.json
ADE20KPart234/ade20k_instance_val_mixed_parts.json
PartImageNet/annotations/train/train_mixed_parts.json
PartImageNet/annotations/train_whole/train_mixed_parts.json
PartImageNet/annotations/val/val_mixed_parts.json
PartImageNet/annotations/val_whole/val_mixed_parts.json
PartImageNet/annotations/test/test_mixed_parts.json
PartImageNet/annotations/test_whole/test_mixed_parts.json

The official Mixed Parts test split is mixed_parts_test.json.

Split Sizes

Split	Samples
Train	2,380,749
Val	1,999
Test	999

Source Data

Download the original source data from the upstream projects:

• ADE20KPart234: download ADE20KPart234.tar.gz from InternRobotics/OV_PARTS.
• PartImageNet: download PartImageNet_Seg.zip from tacju/partimagenet.
• COCO2017 and PACO-LVIS: follow facebookresearch/paco and download paco_lvis_v1.zip. PACO-LVIS annotations reference COCO2017 images, so COCO2017 images are also required.

Data Preparation

From the CALICO repository root, create the default data layout:

mkdir -p data/mixed_parts_data data/coco_2017

Extract ADE20KPart234 and PartImageNet under data/mixed_parts_data:

tar -xzf /path/to/ADE20KPart234.tar.gz -C data/mixed_parts_data
unzip /path/to/PartImageNet_Seg.zip -d data/mixed_parts_data

Extract PACO-LVIS annotations under data/mixed_parts_data/paco_lvis/annotations:

mkdir -p data/mixed_parts_data/paco_lvis/annotations
unzip /path/to/paco_lvis_v1.zip -d data/mixed_parts_data/paco_lvis/annotations

Place COCO2017 images under data/coco_2017:

data/coco_2017/
├── train2017/
└── val2017/

Download this annotation bundle directly into data/mixed_parts_data:

huggingface-cli download PLAN-Lab/MixedParts \
  --repo-type dataset \
  --local-dir data/mixed_parts_data \
  --local-dir-use-symlinks False

If you downloaded the bundle somewhere else, overlay it into data/mixed_parts_data:

cp -r /path/to/MixedParts/* data/mixed_parts_data/

The prepared data should have this structure:

data/
├── coco_2017/
│   ├── train2017/
│   └── val2017/
└── mixed_parts_data/
    ├── mixed_parts_train.json
    ├── mixed_parts_val.json
    ├── mixed_parts_test.json
    ├── ADE20KPart234/
    │   ├── ade20k_instance_train_mixed_parts.json
    │   ├── ade20k_instance_val_mixed_parts.json
    │   ├── images/
    │   └── annotations_detectron2_part/
    ├── PartImageNet/
    │   ├── annotations/
    │   │   ├── train/train_mixed_parts.json
    │   │   ├── train_whole/train_mixed_parts.json
    │   │   ├── val/val_mixed_parts.json
    │   │   ├── val_whole/val_mixed_parts.json
    │   │   ├── test/test_mixed_parts.json
    │   │   └── test_whole/test_mixed_parts.json
    │   └── images/
    └── paco_lvis/
        └── annotations/
            ├── paco_lvis_v1_train.json
            ├── paco_lvis_v1_val.json
            └── paco_lvis_v1_test.json

Evaluation with CALICO

After preparing the data, run evaluation from the CALICO repository root:

python evaluate.py \
  --merged_ckpt_path PLAN-Lab/CALICO \
  --dataset_dir ./data \
  --output_save_path ./evaluate_results/calico_mixed_parts \
  --val_dataset "MixedPartsObjectVal|MixedPartsPartVal" \
  --multi_image_filepath_prefix ./data/mixed_parts_data/mixed_parts_test.json \
  --mode test \
  --compute_metrics

For full setup details, see the CALICO repository documentation.

Licensing and Terms

This repository provides Mixed Parts annotations for research use. The underlying images and source annotations come from ADE20KPart234, PartImageNet, COCO2017, and PACO-LVIS. Users are responsible for following the licenses and terms of the upstream datasets.

Citation

If you use Mixed Parts or CALICO, please cite:

@article{nguyen2025calico,
  title={CALICO: Part-Focused Semantic Co-Segmentation with Large Vision-Language Models},
  author={Nguyen, Kiet A. and Juvekar, Adheesh and Yu, Tianjiao and Wahed, Muntasir and Lourentzou, Ismini},
  journal={In Proceedings for the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2025}
}