PODS: Personal Object Discrimination Suite

🌐Project page            📖Paper            GitHub

We introduce the PODS (Personal Object Discrimination Suite) dataset, a new benchmark for personalized vision tasks.

PODS

The PODS dataset is new a benchmark for personalized vision tasks. It includes:

• 100 common household objects from 5 semantic categories
• 4 tasks (classification, retrieval, segmentation, detection)
• 4 test splits with different distribution shifts.
• 71-201 test images per instance with classification label annotations.
• 12 test images per instance (3 per split) with segmentation annotations.

PODS is split class-wise into a validation set (6 classes per semantic category) and a test set (14 classes per semantic category). All test performance reported in our paper is from the test set of classes.

Within each class, images are divided into a train/retrieval set (3 images) and a test/query set. The test/query set is then further divided into 4 test splits reflecting different distribution shifts.

Metadata is stored in two files:

• pods_info.json:
  • classes: A list of class names
  • class_to_idx: Mapping of each class to an integer id
  • class_to_sc: Mapping of each class to a broad, single-word semantic category
  • class_to_split: Mapping of each class to the val or test split.
• pods_image_annos.json: Maps every image ID to a dictionary:
  • class: The class name that the image belongs to
  • split: One of [train, test] indicating if the image is in the train or test set for that class.
  • test_split: For images in the test split, denotes which distribution-shift test split the image is in: One of [in_distribution, pose, distractors, pose_and_distractors]

Using PODS

Loading the dataset using HuggingFace

To load the dataset using HuggingFace datasets, install the library by pip install datasets

from datasets import load_dataset

pods_dataset = load_dataset("chaenayo/PODS") 

You can also specify a split by:

pods_dataset = load_dataset("chaenayo/PODS", split="train") # or "test" or "test_dense"

Loading the dataset directly

PODS can also be directly downloaded via command:

wget https://data.csail.mit.edu/personal_rep/pods.zip

Citation

If you find our dataset useful, please cite our paper:

@inproceedings{sundaram2025personalized,
  title={Personalized representation from personalized generation},
  author={Sundaram, Shobhita and Chae, Julia and Tian, Yonglong and Beery, Sara and Isola, Phillip},
  booktitle={International Conference on Learning Representations},
  volume={2025},
  pages={92572--92606},
  year={2025}
}