---
language:
- en
tags:
- part_understanding
- segmenting_lmms
- vision-language
pretty_name: Partonomy-Core
---
Repository for the **Partonomy-Core** benchmark introduced in [*PARTONOMY: Large Multimodal Models with Part-Level Visual Understanding*](https://arxiv.org/abs/2505.20759).

Code can be found at [https://github.com/anselblume/partonomy](https://github.com/anselblume/partonomy).

To download the images and evaluation metadata, install the [HuggingFace CLI](https://huggingface.co/docs/huggingface_hub/en/guides/cli) and run 

```bash
hf download partonomy/partonomy-core --repo-type dataset --local-dir .
```

Unpack the images directory with `tar --zst -xf images.tar.zst`.

The `images` directory contains all images used in the **Partonomy-Core** evaluation. Evaluation questions, answer choices, and corresponding parts for each answer choice are contained in `test-qa_pairs.json`. Each question-answer JSON object contains the following fields:
- `image_path` (str): Relative path to the image corresponding to the question.
- `image_label` (str): Label for the object in the image.
- `question` (str): The LMM prompt used for evaluation of parts.
- `question_type` (str): The type of question. These include *identification_with_label* (**identification**), *positive_with_label* (**intersection**), *negative_with_label* (**difference**), *whole_to_part* (**Whole-to-Part**), and *part_to_whole* (**Part-to-Whole**).
- `answer_choices` (list[str]): The multiple choice answer choices to evaluate the model with, using each answer choice individually as the model's response to the question prompt.
- `answer_types` (list[str]): The correctness of each answer choice in the `answer_choices` field.
- `answer_parts` (list[list[str]]): List containing each of the part names sampled for each answer choice.
- `segmentations` (dict): Segmentation masks for each part, where keys are the part names. Segmentation mask objects for each part can be decoded into binary masks using `pycocotools.mask.decode`.
- `object_question` (str): Analogue of `question` LLM prompt, used for evaluation of the object label. Only present for Whole-to-Part and Part-to-Whole questions.
- `object_answer_choices` (list[str]): Answer choices for the `object_question` prompt. Only present for Whole-to-Part and Part-to-Whole questions.
- `object_answer_types` (list[str]): The correctness of each answer choice in the `object_answer_choices` field. Only present for Whole-to-Part and Part-to-Whole questions.
- `object_answer_classes` (list[str]): The object class used for each answer choice in `object_answer_choices`.