Stacking discards camera information from exif data

[mattcaron]

While trying to figure out why my objects' dimensions were all wrong, I checked the exif data and found the following:

Non-stacked:

ExifTool Version Number         : 12.76
File Name                       : scan01_129_fs00.jpg
Directory                       : .
File Size                       : 1448 kB
File Modification Date/Time     : 2026:07:18 19:22:04-04:00
File Access Date/Time           : 2026:07:18 19:22:04-04:00
File Inode Change Date/Time     : 2026:07:19 10:02:17-04:00
File Permissions                : -rw-rw-r--
File Type                       : JPEG
File Type Extension             : jpg
MIME Type                       : image/jpeg
JFIF Version                    : 1.01
Resolution Unit                 : None
X Resolution                    : 1
Y Resolution                    : 1
Exif Byte Order                 : Big-endian (Motorola, MM)
Camera Model Name               : arducam_64mp
Orientation                     : Rotate 270 CW
Software                        : OpenScan3 (Picamera2)
Exposure Time                   : 1/20
ISO                             : 100
Date/Time Original              : 2026:07:18 19:22:04
Image Width                     : 5732
Image Height                    : 3610
Encoding Process                : Baseline DCT, Huffman coding
Bits Per Sample                 : 8
Color Components                : 3
Y Cb Cr Sub Sampling            : YCbCr4:2:0 (2 2)
Image Size                      : 5732x3610
Megapixels                      : 20.7
Shutter Speed                   : 1/20

Stacked:

ExifTool Version Number         : 12.76
File Name                       : stacked_scan01_129.jpg
Directory                       : .
File Size                       : 1620 kB
File Modification Date/Time     : 2026:07:18 21:17:50-04:00
File Access Date/Time           : 2026:07:18 21:56:45-04:00
File Inode Change Date/Time     : 2026:07:18 21:55:13-04:00
File Permissions                : -rw-rw-r--
File Type                       : JPEG
File Type Extension             : jpg
MIME Type                       : image/jpeg
JFIF Version                    : 1.01
Resolution Unit                 : None
X Resolution                    : 1
Y Resolution                    : 1
Image Width                     : 3610
Image Height                    : 5732
Encoding Process                : Baseline DCT, Huffman coding
Bits Per Sample                 : 8
Color Components                : 3
Y Cb Cr Sub Sampling            : YCbCr4:2:0 (2 2)
Image Size                      : 3610x5732
Megapixels                      : 20.7

I don't know if this is actually an important issue or not, because using the original images didn't fix the object size either.

(If anyone knows what I need to do to get Meshroom to reconstruct the object with the correct dimensions, I'd appreciate a note telling me how. Until then, I'll keep digging. Cheers!)

[bgrupczy]

Not sure on this either but great potential find!

[esto-openscan]

The focus stacking task ignores the previously injected exif data and could be enhanced to use it when creating the resulting image. However, we would need to consider which exif fields are meaningfull, because the resulting images may have odd lens properties. For example we can't derive the camera pose from the stacking process itself. Maybe a virtual OpenScan3 Stack Camera would be a great addition to distinguish them later and choose proper settings in the reconstruction pipeline.

Arbitrary object size and rotation are an inherent problem in photogrammetry pipelines. One elegant solution could be using the camera poses from the pictures as priors. Meshroom seems to have an issue importing camera poses, which is closed as stale but may not be resolved: alicevision/AliceVision#1748 Another thing to watch out is the fragility of the pipeline if you try to reconstruct from more than one scan with a flipped object. Most likely it will break because the coordinate systems from different scans won't fit together.

Another widely used solution for object size and scaling would be using markers (CCtags, see youtube video in linked issue). Another possibility is to use a known reference object for scaling the entire scene as a post processing step. If your object allows it you may even use the distance between two easy measurable points and scale the mesh accordingly after reconstruction.