Photogrammetry Optimization
33 views
Skip to first unread message
Seemal Tahir
Dec 20, 2025, 7:32:08 AM12/20/25
to AliceVision
Hi,
I was wondering if there's a certain set of instructions or standards for taking images that indicate what would produce the best possible reconstruction? I'm looking for the best possible quality in my final reconstruction after running the images through Meshroom. Thanks!
Iris
Jun 5, 2026, 4:13:40 AM (9 days ago) Jun 5
to AliceVision
Depending on the camera you use, there are different factors at play.
Meshroom's documentaiton recommends that when using a DSLR, you are advised to use fixed focal point lenses (lenses that can't zoom, also called prime lenses).
If you do use a zoomlens, use the wide-angle side and the tele-angle side only for the best results.
The higher the teleside focal point (the bigger amount of mm's), the more important choosing a smaller aperture opening becomes to make sure as much is sharp in the resulting picture.
All unsharp elements will either be ignored, or can screw up the calclulation.
Another thing to make sure is to prevent motion blur, harsh shadows and changing light conditions between pictures. A monopod can help prevent motion blur.
Something else to avoid: the use of the pop-up flash or an external flash device.
This is because the shadows it creates are not only harsh, but they also come from the camera's perspective, which can mess with depth and object detection.
Try to fill the frame with the object as much as possible. Depending on the style of photography (turntable or moving around because the object can not be put on a turntable), work in as much of an accurate curve around the object as possible as well. Take pictures close by, further away, and even further away.
That way the software can "watch" what is underneath a piece of the object that from far away might be difficult to view. It get's even more obvious for things that are obscured when far away but you can "look behind" when closer by.
If working with a DSLR, try to feed the RAW's. I have noticed that any software that works on the images during conversion to JPEG and corrects lens distortion can decrease the accuracy of the model.
Sometimes it can even cause some camera's to be excluded from the point cloud generation (I literally tested this with pictures from a trashcan: the JPG's saw 3 exclusions and a very messy, incorrect point cloud, while the RAW's worked flawless).
This is where choosing good prime lenses can help as well: I recently acquired a second hand 14mm Canon lens (the widest non-fisheye) that is practically distortionless, apart from the exaggerated size difference between the fore- and background.
But probably the most important thing: make sure anything behind the object in terms of the background suffers as little distortion as possible.
While the DepthMap and DepthMapFilter node can be used to exclude more parts from the background by changing the Max View Angle number, to much distortion can again, mess with perspective detection.
I changed the default number for the Max View Angle setting, which I believe as the "from installation" default is 70, and set it to 45.
While some things still apply when using a digital point-and-shoot camera, or even a cellphone, quite a lot of the calculation is done by deriving information about the camera and lens used from ExIf information.
Unfortunately not all cell phones produce correct ExIf, and in the case of lens info, all kinds of weird stuff can happen due to how tiny the sensor is.
Some report the 35mm equivalent focal length only (what the lens on the phone looks like on a full-frame DSLR, with the sensor size being that of a 35mm analog film frame size from back in the day), some list the phone's focal length.
(I had trouble with this kind of thing in a different program called fSPY, that is not about photogrammetry, but about placing a 3D camera in the 3D space and calculating the position (X, Y and Z) of said camera based on a picture where you draw converging lines over known axes.)
Meshroom's documentaiton recommends that when using a DSLR, you are advised to use fixed focal point lenses (lenses that can't zoom, also called prime lenses).
If you do use a zoomlens, use the wide-angle side and the tele-angle side only for the best results.
The higher the teleside focal point (the bigger amount of mm's), the more important choosing a smaller aperture opening becomes to make sure as much is sharp in the resulting picture.
All unsharp elements will either be ignored, or can screw up the calclulation.
Another thing to make sure is to prevent motion blur, harsh shadows and changing light conditions between pictures. A monopod can help prevent motion blur.
Something else to avoid: the use of the pop-up flash or an external flash device.
This is because the shadows it creates are not only harsh, but they also come from the camera's perspective, which can mess with depth and object detection.
Try to fill the frame with the object as much as possible. Depending on the style of photography (turntable or moving around because the object can not be put on a turntable), work in as much of an accurate curve around the object as possible as well. Take pictures close by, further away, and even further away.
That way the software can "watch" what is underneath a piece of the object that from far away might be difficult to view. It get's even more obvious for things that are obscured when far away but you can "look behind" when closer by.
If working with a DSLR, try to feed the RAW's. I have noticed that any software that works on the images during conversion to JPEG and corrects lens distortion can decrease the accuracy of the model.
Sometimes it can even cause some camera's to be excluded from the point cloud generation (I literally tested this with pictures from a trashcan: the JPG's saw 3 exclusions and a very messy, incorrect point cloud, while the RAW's worked flawless).
This is where choosing good prime lenses can help as well: I recently acquired a second hand 14mm Canon lens (the widest non-fisheye) that is practically distortionless, apart from the exaggerated size difference between the fore- and background.
But probably the most important thing: make sure anything behind the object in terms of the background suffers as little distortion as possible.
While the DepthMap and DepthMapFilter node can be used to exclude more parts from the background by changing the Max View Angle number, to much distortion can again, mess with perspective detection.
I changed the default number for the Max View Angle setting, which I believe as the "from installation" default is 70, and set it to 45.
While some things still apply when using a digital point-and-shoot camera, or even a cellphone, quite a lot of the calculation is done by deriving information about the camera and lens used from ExIf information.
Unfortunately not all cell phones produce correct ExIf, and in the case of lens info, all kinds of weird stuff can happen due to how tiny the sensor is.
Some report the 35mm equivalent focal length only (what the lens on the phone looks like on a full-frame DSLR, with the sensor size being that of a 35mm analog film frame size from back in the day), some list the phone's focal length.
(I had trouble with this kind of thing in a different program called fSPY, that is not about photogrammetry, but about placing a 3D camera in the 3D space and calculating the position (X, Y and Z) of said camera based on a picture where you draw converging lines over known axes.)
Op zaterdag 20 december 2025 om 12:32:08 UTC schreef Seemal Tahir:
Reply all
Reply to author
Forward
0 new messages