Holopix50k: A Large-Scale In-the-wild Stereo Image Dataset

Holopix50k: A Large-Scale In-the-wild Stereo Image Dataset

best computer vision stereo dataset Holopix50k

A team of Computer vision researchers at Leia Inc. have released a large-scale database of 49,368 (~50k) stereoscopic image pairs collected from the popular Lightfield social media app Holopix™. This is the largest dataset of stereoscopic image pairs ever released publicly that contain in-the-wild scenarios captured from mobile phones. For context, the second-largest dataset in this category consists of only 1024 stereoscopic image pairs, almost 50 times less! The dataset is available for download immediately on the project page:

In the associated paper, that can be found at, the researchers go through how they collected the dataset and also explore the nature of the data present. The dataset contains about 21k humans, followed by a large number of animals, vehicles, and furniture. This shows the diverse nature of scenarios this dataset entails, with both indoor and outdoor scenes. Additionally, the researchers point out interesting categories of images the dataset contains such as fire, toys, selfies, parties, and snow. All of this is possible due to the Holopix™ platform containing a large number of images uploaded by users of the platform worldwide from their LitByLeia® Lightfield mobile devices. 

Holopix50k dataset images

The Holopix50k dataset has already shown improvements to state-of-the-art results for cutting edge computer vision tasks such as stereo super-resolution (which uses the information present in two images in order to super-resolve one of them), and self-supervised monocular depth estimation (which aims to predict the depth from a single image, by training on the depth calculated from a pair of images). Additionally, in their paper, the team at Leia Inc. has shown a variety of practical applications that already use the Holopix50k dataset, such as the disparity estimation neural networks that currently run on Holopix.

Holopix50k dataset benchmark computer vision

The Holopix50k dataset fills an important gap in the stereoscopic computer vision research community. Most of the current popular stereo datasets such as KITTI have a limited number of scenes and are specific to a domain, such as self-driving cars. Other datasets that might be larger are still not large enough to enable generalizability to mobile photography scenarios (due to increasing number of dual-cameras). The Holopix50k dataset being magnitudes larger, and having high quality stereo images, which are from what the researchers call “in-the-wild” scenes, puts it in a good spot to become a popular dataset in this space. The researchers behind this dataset mention in their paper that they hope for other researchers to create new and exciting techniques based on in-the-wild stereo images in this dataset. We hope so too, and can’t wait to see what comes out of this work!

For more details visit any of the links below

  • Authors: Yiwen Hua •  Puneet Kohli •  Pritish Uplavikar •  Anand Ravi •  Saravana Gunaseelan •  Jason Orozco •  Edward Li
  • Paper: 
  • Dataset: Holopix50k
  • Published: 25 March 2020

For More Resources: