Using the new approach, researchers have launched a groundbreaking geolocation technique. This radical advancement greatly increases both the speed and effectiveness of identifying locations based on images. This entirely new mechanism takes advantage of the latest and greatest artificial intelligence and machine learning techniques. It was just published in the IEEE Transactions on Geoscience and Remote Sensing. Mapillary’s computer vision technology will completely change how we analyze images to understand geographic context. Its use is multifaceted, from military tactics to arranging family pictures.
The geolocation method’s high success rate is an impressive 97 percent in the first stage of directing users to specific locations and more than compensates for this disadvantage. Such success continues to hold even with images using 180-degree FOV. When it comes to pinpointing a specific site, the accuracy drops but is still quite strong at 82 percent. Unlike current models, this pioneering method gets to results in a fraction of the time. It’s much more memory efficient, requiring just 35 megabytes. By comparison, the second most compact model takes up 104 megabytes. This new technique is more than twice as fast and has a much lower memory footprint.
This cutting-edge method matches the code from a single ground-level photo. It then overlays that code on aerial images saved to the database. It matches each candidate to the five most similar candidates for possible better matches, enabling quick comparison of location even without the use of metadata. This feature might be especially useful in a defense context, allowing quick geolocation of photos taken in the field.
“We train the AI to ignore the superficial differences in perspective and focus on extracting the same ‘key landmarks’ from both views, converting them into a simple, shared language,” said Peng Ren, one of the researchers involved in the development of this method.
The possible uses for this technology are enormous. In addition to military uses, the method could assist emergency responders in quickly locating disaster sites using photographs taken during crises. In the short term, over the next five years, it should allow automatic geotagging of archival family photos. This will enable individuals to easily document and retrieve their unique life journeys.
While there is considerable excitement about this development, some in the industry don’t consider it to be groundbreaking. Nathan Jacobs, who runs a competing startup called Orpheo, said, “I don’t think that this is a particularly groundbreaking paper.” He cautioned that though the advancements are impressive, they might not be fully moving away from the old approach.
Li also admitted that although this whole approach isn’t entirely new, this ‘approach … marks a significant improvement within the field’. The researchers’ emphasis on improving upon currently used techniques is a great sign to see in terms of advancing geolocation techniques.