README.md

---
license: mit
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
  - split: test_dense
    path: data/test_dense-*
dataset_info:
  features:
  - name: image
    dtype: image
  - name: mask
    dtype: image
  - name: label
    dtype: string
  - name: scene_type
    dtype: string
  splits:
  - name: train
    num_bytes: 27536486.0
    num_examples: 300
  - name: test
    num_bytes: 1057988392.0
    num_examples: 10888
  - name: test_dense
    num_bytes: 142893072.0
    num_examples: 1200
  download_size: 1168238221
  dataset_size: 1228417950.0
tags:
- personalization
- instance_detection
- instance_classification
- instance_segmentation
- instance_retrieval
---
# PODS: Personal Object Discrimination Suite
<h3 align="center"><a href="https://personalized-rep.github.io" style="color: #2088FF;">🌐Project page</a>&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp
  <a href="https://arxiv.org/abs/2412.16156" style="color: #2088FF;">📖Paper</a>&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp
<a href="https://github.com/ssundaram21/personalized-rep" style="color: #2088FF;">GitHub</a><br></h3>

We introduce the PODS (Personal Object Discrimination Suite) dataset, a new benchmark for personalized vision tasks. 
<p align="center">
  <img src="https://cdn-uploads.huggingface.co/production/uploads/65f9d4100f717eb3e67556df/uMgazSWsxjqEa4wXSmkVi.jpeg" alt="pods.jpg" />
</p>


## 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:
```
@article{sundaram2024personalized,
  title   = {Personalized Representation from Personalized Generation}
  author  = {Sundaram, Shobhita and Chae, Julia and Tian, Yonglong and Beery, Sara and Isola, Phillip},
  journal = {Arxiv},
  year    = {2024},
}
```