Elith Inc. announces that its employees achieved top rankings in the data analysis competition "#23 Turing × atmaCup 2nd" (held January 16–25, 2026), organized by atma Inc. and co-hosted by Turing Inc.
This competition was open to the public, with 282 teams participating and 2,216 prediction submissions made, making it a highly competitive event. Among them, our employees achieved the following results.
■ Results
- 5th place: Nagasawa
- 11th place: Fujihara
Both are top-tier results achieved among a large field of participants, demonstrating the high level of technical skill Elith has in the fields of data analysis and machine learning.
Elith places importance on improving practical skills in the AI and data science fields, and actively promotes technical development through external competitions in addition to day-to-day project work.
Through such efforts, we will continue to contribute to the real-world deployment of cutting-edge technology and to delivering value to our clients.
■ What is atmaCup
atmaCup is a data analysis competition hosted by atma Inc.
It is characterized by a format in which participants build analytical and predictive models within a limited time based on data provided by the organizer, competing on accuracy, and it is known as a high-level competition that attracts many practitioners and researchers in the data science field.
The 23rd competition was held jointly with Turing Inc.
In this competition, camera images captured from real driving scenes, anonymized masked-region images, and location data (latitude/longitude) were provided. Based on this data, participants built location-estimation models for driving images with unknown location information, and competed on accuracy.
■ Comments
Nagasawa (AI Engineer)
In this competition, the dominant approach was to search a database for images similar to the given image and estimate location based on their location data. However, this method alone cannot pinpoint a completely accurate location. To address this, we corrected the location as post-processing using a technique that reconstructs 3D structure from images. Furthermore, by focusing on the fact that the data came from an onboard vehicle camera and taking advantage of the fact that roads are mostly straight to constrain the direction of correction, we were able to greatly improve accuracy. Going forward, I would like to actively apply this kind of technology to real-world projects as well. My heartfelt thanks to the organizers for hosting this event.
Fujihara (AI Engineer)
I broadly experimented with the latest methods in image retrieval (Visual Place Recognition). What was particularly striking was that CNN-based models (EigenPlaces) and ViT-based models (SALAD/BoQ/AnyLoc/EDTformer) returned surprisingly different retrieval results. Taking advantage of this orthogonality, I built a pipeline that combines and PCA-compresses the features of six models, then computes coordinates via a distance-constrained weighted average. Achieving stable accuracy through an approach of retrieval plus statistical post-processing, rather than an end-to-end regression model, was a major learning experience. My thanks to the organizers for such a wonderful opportunity.