Google has released Google-Landmarks-v2, an improved dataset for Landmark Recognition & Retrieval. The Google Landmarks V2 release is the second iteration of the Google-Landmarks dataset previously released in March 2018. This new version contains 5 million images of more than 200,000 different landmarks. The images were collected from photographers around the world who labeled their own photos and supplemented them with historical and lesser-known images from Wikimedia Commons.
To accompany the release, Google has open-sourced DELF, the Tensorflow-based code and associated model for large-scale instance-level image recognition. DELF leverages Detect-to-Retrieve, an image representation suited for the retrieval of specific object instances.
Two companion Kaggle challenges have also been launched: Landmark Recognition 2019 and Landmark Retrieval 2019. The recognition challenge aims at the comprehensive identification of landmarks, while the goal of the retrieval challenge is to find all the similar representations of a particular landmark among a large collection of images. Adding to the inherent difficulties of large scale image recognition is the presence of "junk" images without any landmarks, and the scarcity of certain landmarks in the training dataset.
The winning teams will be invited to present their methods at the Second Landmark Recognition Workshop at the 2019 Conference on Computer Vision and Pattern Recognition in Long Beach, California, later this year. Last year, the Kaggle challenges based on the initial Google Landmark dataset attracted over 500 teams of data scientists.
Landmark recognition differs from other image recognition problems in three ways:
- It is an instance-level recognition problem: instead of recognizing general entities such as buildings or mountains, the goal is to recognize specific monuments or landmarks such as the Eiffel tower or Niagara Falls.
- The number of different entities is much larger than classic image recognition challenges; Image-Net ILSVRC challenge and some landmarks have scarce representations.
- Landmarks are static objects that rarely change. The variations in their images arise due to image capture conditions, such as exposure or viewpoint. This is different from general image recognition, where objects (dogs, cars, ...) have many variations.
Landmark recognition is used for augmented reality mobile applications that can recognize a captured landmark and retrieve related information. Google already enables Landmark recognition in mobile phones via its Firebase ML-kit API, which offers a dedicated landmark identification feature. Other companies such as Blippar also develop real-world object recognition applications that include Landmark recognition.
Google's first attempts at landmark recognition dates back to 2009, with the development of a landmark recognition engine which was already deemed to be 80% accurate at the time.
