Office-Home

Overview

Office-Home is a domain adaptation and domain generalization benchmark with 15,588 images of 65 everyday objects found in offices and homes (alarm clock, backpack, laptop, mug, …). Every class appears in four domains. Compared to PACS it has many more classes and few images per (domain, class) pair, which makes it a harder fine-grained transfer task.

Domain	Content	Images
art	Paintings, sketches and artistic depictions	2,427
clipart	Clip art images	4,365
product	Product photos without background	4,439
real_world	Regular camera photos	4,357
Total		15,588

Dataset Configurations

The original release (OfficeHomeDataset_10072016.zip) has no official train/test split, so the builder offers two protocols.

random (default)

A stratified 80/20 train/test split inside every (domain, class) pair, so every domain and class appears in both splits. Images are assigned by a SHA-256 hash of their bytes, so the split is reproducible and does not depend on file order. Use this config as an in-distribution benchmark.

art, clipart, product, real_world (leave-one-domain-out)

train holds every image from the other three domains, test holds every image from the named domain. This is the standard domain generalization protocol: the model never sees the test domain during training.

Data Structure

When accessing an example using ds[i], you will receive a dictionary with the following keys:

Key	Type	Description
image	PIL.Image.Image	Variable resolution RGB image (non-RGB inputs are converted)
label	int	Class label (0-64)
domain	int	Domain index (0: art, 1: clipart, 2: product, 3: real_world)

Usage Example

Basic Usage (random split)

from stable_datasets.images import OfficeHome

ds_train = OfficeHome(split="train")
ds_test = OfficeHome(split="test")

sample = ds_train[0]
print(sample.keys())  # {"image", "label", "domain"}
print(ds_train.features["label"].names[sample["label"]])

ds_torch = ds_train.with_format("torch")

Leave-One-Domain-Out

from stable_datasets.images import OfficeHome

# Train on art, product and real_world; evaluate on clipart
ds_train = OfficeHome(config_name="clipart", split="train")
ds_test = OfficeHome(config_name="clipart", split="test")
print(len(ds_train), len(ds_test))  # 11223 4365

References

• Official website and download: https://www.hemanthdv.org/officeHomeDataset.html

• Paper: https://arxiv.org/abs/1706.07522

• License: fair-use notice, non-commercial research and educational use only (see the official website)

Citation

@inproceedings{venkateswara2017deep,
  title     = {Deep Hashing Network for Unsupervised Domain Adaptation},
  author    = {Venkateswara, Hemanth and Eusebio, Jose and Chakraborty, Shayok and Panchanathan, Sethuraman},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {5018--5027},
  year      = {2017}
}