Dataset description

Mosaic REALMAP Demo dataset with 9114 images collected with panoramic camera with global shutter sensors. You can find more about full dataset in this article: Mosaic Prague REALMAP Each image has its position written in EXIF metadata populated from IMU device paired with the camera.

Images from each sensor are electrically syncronized up to several nanoseconds and the whole device can be considered as a multicamera rig. Additional camera information, including .gpx track is provided.

Camera calibration is provided in COLMAP compatible format.

Dataset

All extracted data is anonymized to comply with GDRP requirements. Faces and car plates are blurred and blurring masks are provided also.

Each image has its positional and time EXIF metadata build from frame_times.csv and route.gpx (alternatively, just route_frame_sycned.gpx can be used). Tools like COLMAP recognize this data and it may save some time for matching these images.

DEMO_full_data

The folder contains original images from each camera sensor extracted from video stream it writes. 8-bit images, with fisheye lenses. Be careful using this data with tools which do not support fisheye type of distortion like Reality Capture. For such applications undistorted data is prefered. It contains calibration in colmap format inside calib_colmap.txt which is compatible with cameras.txt output from COLMAP.

COLMAP

Project

Import reel folder reel_0017_20240117-094718 with cam0, cam1, ..., cam5 sub-folders inside. Select new database.

Feature extractor

Use "shared per sub-folder" option. Use OPENCV_FISHEYE model, but parameters can be modified later. With this option turned on images in each camN folder will share intrinstics and distortion parameters as they should. As a mask_path select reel_0017_20240117-094718_masks folder. After features have beend extracted, open database and modify camera parameters in the database using calib_colmap.txt file.

Feature matching

Exhausive matching may take significant amout of time. To avoid that use combination of sequential, spatial, vocabulary and transitive matching (or just run them in this order).

Reconstruction

It may help to switch off optimization of camera focal length, distortion parameters and principal point (principal point refinement is switch off by default) to speed up reconstruction a bit.

Post reconstruction step

Using COLMAP's CLI run rig_bundler_adjuster subcommand to enforce camera rig constrains.

Small colmap project

In folder DEMO_small_subset_and_colmap_prj there is a small COLMAP project example, including .db, project.ini and exported model. It also includes depth map project with extracted dense point cloud. This may be useful if you wander how images should be loaded and db organized and matched.

DEMO_full_data_undistorted

The same as DEMO_full_data except that images and masks are unfisheyed and undistorted. A simple distortion model can be applied to them, see calib_colmap_undistorted.txt file for their parameters.

Reality Capture

For reality capture you can initialize all cameras as the same image group so they share the same distortion and intrinsics. Focal length can be initilized from values in calib_reality_capture.txt

TODO

Explain calibration format, share full .json calibration and visualization script
Provide imu2camera calibration, provided poses for all images