Releasing first of a kind large-scale localization datasets in crowded indoor spaces

Visual localization algorithms estimate the camera pose of a given image which is an important task in robotics, autonomous driving and in augmented reality (AR) and the results of modern visual localization methods have been promising enough for some real-world applications to have been developed such as AR navigation.

Challenges. Indoor environments differ from outdoor scenes in a number of important aspects concerning visual localization. For example, although lighting conditions in indoor spaces tend to be more constant than those in the outside world, indoor spaces are typically narrower so small camera movements can cause large changes in viewpoint. Dynamic objects like people, trolleys, TV screens, areas with no texture and repetitive patterns are also challenging for visual localization and there are many more of these in indoor spaces.

Existing datasets. In outdoor environments, the structure-from-motion (SfM) technique is effective in generating ground truth camera poses, which are necessary for all localization algorithms. Unfortunately, due to the aforementioned challenges, large-scale SfM is more difficult and often impossible to apply to indoor spaces. As a consequence, the area covered by existing indoor datasets (Stanford, ScanNet, Matterport 3D, 7 scenes, InLoc and Baidu) is only sparsely sampled by the provided images (e.g. only one image every 2 meters).

Furthermore, while the presence of people is an important aspect for a realistic dataset, privacy issues and limitations of data acquisition (standard SfM often fails) typically hinder the inclusion of crowds in state-of-the-art datasets.

NAVER LABS datasets. To contribute to the need for more training data, and in particular data of indoor environments, NAVER LABS has made available five new localization datasets acquired in challenging real-world indoor environments. The datasets were captured in a large shopping mall and a busy metro station in Seoul, South Korea and include images and depth maps obtained from a dedicated mapping platform with ten cameras and two 3D laser scanners. Compared to existing indoor datasets, the proposed datasets provide dense image sampling and cover many challenges of indoor visual localization as shown in Figure 1. In addition, the datasets consist of captures of sequences with time intervals of up to 128 days providing changes over time. To address the problem that SfM cannot be used alone to estimate precise ground truth camera poses for large indoor spaces, we developed a novel, automated pipeline that utilizes both LiDAR SLAM and SfM algorithms. The proposed pipeline applies LiDAR-based pose-graph SLAM to estimate the trajectories of our dedicated mapping platform M1X. For full details on the datasets see our CVPR 2021 paper (4) and/or article Indoor Localization Dataset V1.0 (Dept. & Metro St.)

Kapture. The datasets are provided in the unified data format kapture. Kapture is being increasingly adopted as the toolbox that accompanies it makes the execution of modern visual localization methods easy and the conversion into other formats such as OpenMVG and COLMAP simple.

Evaluation. The datasets are part of the Long-Term Visual Localization Benchmark. Localization results of your own algorithm on the test sets can be evaluated and compared against other methods. For each dataset (Gangnam Station and Hyundai Department Store), the results have to be submitted as one text file containing all floors (2 for Gangnam Station, 3 for Hyundai Department Store).
Text file format:

# image_path qw qx qy qz tx ty tz 
2019-08-21_11-47-29/galaxy/AC01324968_1566355938331000.jpg -0.238452 -0.247522 -0.618811 0.706361 670.685 -16.5772 -51.8993
2019-08-21_11-47-29/galaxy/AC01324968_1566356238363000.jpg 0.701688 0.528133 -0.285814 0.383432 -634.116 -26.8537 -180.42
2019-08-21_11-47-29/galaxy/AC01324968_1566356280347000.jpg 0.464522 0.351865 -0.514493 0.629052 -239.154 -97.3573 -580.324
...

• one line per image
• image_path is the relative path of an image with respect to the records_data folder
• qw qx qy qz represent the rotation from world to camera coordinates as a unit quaternion (as in kapture)
• tx ty tz is the camera translation, not the camera position (as in kapture)

Note that there is a script in kapture (to be used with the full_file_name option) that exports the kapture trajectory to this format.

(1) Dataset: https://www.naverlabs.com/en/datasets

(2) kapture data format: https://github.com/naver/kapture

(3) kapture toolbox (visual localization): https://github.com/naver/kapture-localization

(4) Large-scale localization datasets in crowded indoor spaces, Donghwan Lee, Soohyun Ryu, Suyong Yeon, Yonghan Lee, Deokhwa Kim, Cheolho Han, Yohann Cabon, Philippe Weinzaepfel, Nicolas Guerin, Gabriela Csurka, and Martin Humenberger, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. [arXiV 2105.08941]