# SUN397

SUN397 is a benchmark dataset for scene recognition, containing 108,754 images across 397 scene categories, widely used to evaluate computer vision algorithms.

SUN397 is a large-scale dataset for scene recognition, introduced in 2010 by researchers at the [MIT Computer Science and Artificial Intelligence Laboratory](https://www.wikiprompt.org/wiki/mit-csail) and Princeton University. It contains 108,754 images distributed across 397 distinct scene categories, making it one of the most comprehensive resources for evaluating algorithms that classify images by the type of environment they depict, such as a bedroom, a ski resort, or a library. The dataset was designed to push the boundaries of scene understanding beyond object-centric tasks, emphasizing the holistic context of an image.

The dataset is organized hierarchically, with categories grouped into broader classes like indoor, outdoor natural, and outdoor man-made. Each category contains at least 100 images, ensuring sufficient samples for training and testing. The images are sourced from public collections, including Flickr and other online repositories, and are manually curated to ensure relevance and quality. SUN397 is commonly used in conjunction with the SUN benchmark, which also includes attributes and object annotations for a subset of images.

## Construction and Annotation

The creation of SUN397 involved a rigorous process to ensure diversity and coverage. The researchers first compiled a list of scene categories from various sources, including dictionaries and existing datasets, then expanded it to 397 categories. Images were collected via web searches and user submissions, then filtered to remove duplicates and irrelevant content. Each image was manually verified to belong to its assigned category. The dataset was split into training and testing sets, with a standard protocol of using 50 images per category for training and 50 for testing, though other splits are also used.

## Role in Scene Recognition Research

SUN397 has become a standard benchmark in [computer vision](https://www.wikiprompt.org/wiki/computer-vision) and [machine learning](https://www.wikiprompt.org/wiki/machine-learning). It is frequently used to evaluate [convolutional neural networks](https://www.wikiprompt.org/wiki/convolutional-neural-network) and other [deep learning](https://www.wikiprompt.org/wiki/deep-learning) models. The dataset challenges algorithms to distinguish between visually similar scenes, such as different types of kitchens or libraries, and to generalize across diverse lighting, perspectives, and occlusions. Many state-of-the-art models report performance on SUN397 as a key metric, often achieving over 90% accuracy on the test set.

## Comparison with Other Datasets

SUN397 complements other popular datasets like [Places365](https://www.wikiprompt.org/wiki/places365) and [ImageNet](https://www.wikiprompt.org/wiki/imagenet). While ImageNet focuses on object classification, SUN397 emphasizes scene-level understanding. Places365, introduced later, is a larger scene dataset with 365 categories, but SUN397 remains valuable due to its finer-grained categories and its role in early scene recognition research. The dataset also includes a subset with attribute labels, enabling tasks like scene attribute prediction.

## Impact and Legacy

The introduction of SUN397 spurred significant advances in scene recognition, contributing to the development of models that understand context and layout. It has been used in applications ranging from autonomous driving to content-based image retrieval. The dataset is publicly available for research purposes and is widely cited in academic literature. Its hierarchical structure and clear evaluation protocols have made it a durable resource, still referenced in contemporary studies as of 2025.

## See Also

- [Places365](https://www.wikiprompt.org/wiki/places365)
- ImageNet
- [Computer vision](https://www.wikiprompt.org/wiki/computer-vision)
- [Convolutional neural network](https://www.wikiprompt.org/wiki/convolutional-neural-network)

---
Source: https://www.wikiprompt.org/wiki/sun397
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-13T03:53:08.01777+00:00