Digging Deeper: Learning Multi-Level Concept Hierarchies
Paper • 2603.10084 • Published
Datasets:
Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.12/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 82, in _split_generators
raise ValueError(
ValueError: The TAR archives of the dataset should be in WebDataset format, but the files in the archive don't share the same prefix or the same types.
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 65, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Dataset Card for PseudoKitchens
Dataset Summary
PseudoKitchens is a synthetic dataset of photorealistic 3D kitchen renders with ground-truth concept (object) annotations. It is designed for tasks such as recipe classification and spatial concept localisation. The dataset consists of kitchen scenes containing ingredients for recipe classification tasks, with each scene accompanied by annotations that describe the location of every ingredient in it.
PseudoKitchens-2 is a version of the PseudoKitchens dataset designed to enable the discovery and evaluation of two-level concept hierarchies by distinguishing between variations of the same ingredient (e.g., "red apples" vs "green apples").
Our repository contains code for loading the dataset.
Dataset Structure
Data Instances
A typical instance in the dataset consists of a rendered image, a corresponding JSON metadata file, and segmentation masks in the Cryptomatte format.
Data Splits
Both PseudoKitchens and PseudoKitchens-2 have three splits.
| Split | Number of Images |
|---|---|
| train | 10,000 |
| validation | 1,000 |
| test | 1,000 |
Dataset Creation
PseudoKitchens is generated using Blender 4.5, a professional open source 3D graphics software package. Our approach leverages physically-based rendering to create photorealistic images whilst maintaining complete experimental control over scene properties.
Kitchen environments are constructed using 3D assets sourced from BlenderKit, all licensed under Royalty Free or Creative Commons CC0 licences. The base kitchen layouts feature countertops, cabinets, appliances, and storage areas. We manually curated five distinct kitchen layouts.
You can use the scripts in this repository to generate your own custom PseudoKitchens datasets. These scripts were developed using Blender 4.5 LTS.
Instance Generation
For each generated image:
A kitchen layout, floor and wall textures are selected uniformly at random. The light position, intensity and colour temperature are chosen randomly.
The camera viewpoint is randomised within predefined bounds for each kitchen, varying both angle and distance to ensure diverse perspectives whilst maintaining ingredient visibility. In some cases, not all of the ingredients placed will be visible. This mirrors real images, where some features might be occluded or out of the shot. The spatial concept annotations describe exactly which ingredients are visible in each image.
A physics-aware placement system positions ingredients on available surfaces (countertops or tables) using weighted random selection based on surface area. Objects are not allowed to overlap, and are randomly rotated and scaled to provide variation. Task-irrelevant objects, such as saucepans and cooking utensils, are randomly placed in scenes.
Annotations
Annotation process
Annotations are automatically generated during the rendering process. We provide:
- Recipe Labels for recipe classification tasks.
- Concept Location Annotations: Pixel-perfect segmentation masks for every ingredient using Blender's Cryptomatte support.
- Instance Information: A JSON file containing all the information needed to recreate the scene. This includes the names of all the objects in the image, along with complete scene parameters including camera position, lighting conditions, material assignments, and object transformations. Note that some objects in the JSON file may be subject to occlusion or outside the viewing volume and so may not be visible in the image. Use the segmentation masks in the
.exrfiles (which use the Cryptomatte format) to determine which objects are actually visible.
Personal and Sensitive Information
The dataset does not contain any personal or sensitive information.
Additional Information
Licensing Information
The datasets, and the scripts used to create them, are licensed under the MIT license.
Citation Information
If you use PseudoKitchens or generate your own images/dataset, please cite:
@inproceedings{hill2026hierarchical,
title={Hierarchical Concept-based Interpretable Models},
author={Oscar Hill and Mateo Espinosa Zarlenga and Mateja Jamnik},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://arxiv.org/abs/2602.23947}
}
If you use PseudoKitchens-2, please cite:
@inproceedings{hill2026digging,
title={Digging Deeper: Learning Multi-Level Concept Hierarchies},
author={Oscar Hill and Mateo Espinosa Zarlenga and Mateja Jamnik},
booktitle={ICLR 2026 Workshop on Principled Design for Trustworthy AI - Interpretability, Robustness, and Safety across Modalities},
year={2026},
url={https://arxiv.org/abs/2603.10084}
}
- Downloads last month
- -