3D reconstruction of camera rig with Colmap

Замовник: AI | Опубліковано: 15.02.2026

Our project is to produce 6DoF films using a 46-camera rig. We have a hemisphere-shaped rig with 46 cameras (currently, we're only using 9 for testing). The cameras are distributed around the hemisphere and face outwards. This is important for the Colmap calculation. There are many overlaps in the images, but not between all cameras. The cameras on the far left and far right, for example, don't overlap. We built the rig in a 3D program and positioned the cameras quite precisely. We then exported the camera data to JSON using a script, which we then imported into Colmap as the start poses. We spent a long time developing a script that correctly exports the poses and imports and uses them correctly in Colmap. I'm happy to make this script available. Also, the JSON containing the start extrinsics. We also calibrated each individual camera with OpenCV and have very precise intrinsics for each camera. This is also read and used by Colmap. I will also provide the JSON. Then we made a test recording with our rig. It's just over a minute long at 30fps and 4k resolution. The rig itself is stationary. We moved and panned the rig in the recording. This results in a lot of parallax and reveals many details in the images. I will also provide the videos. In the end, we have a precise target format that we need. The poses need to be refined. Currently, they are rough, but in the end, they must be calculated very accurately. We then need two files: a JSON containing the poses and intrinsics, and the calculated point cloud. Afterward, we work with the software 4D Gaussian Studio v0.1 to calculate 6DoF videos using this data and our video footage. The problem is that the software doesn't calculate the poses correctly, so we first have to do this in Colmap and then import the data into the software. I'd be happy to provide a template of these two files. We'll then need the Colmap script from you, into which we can input our extrinsic and intrinsic data and video files. It will then calculate the poses and the point cloud, which we'll then integrate into our workflow. Colmap always messes up the rig and doesn't produce a clean calculation. Theoretically, we could use other software, but I think Colmap is the best option, right?