Losing majority of photos on orientation. Crashing on occasion during orientation

[Dangerfro]

Hello, I've been messing around with 3df zephyr and decided to try it with some photos I found online of a 28mm scale miniature. However, despite me trying a myriad of settings it usually ever captures the last 12 images (as it is a set from the top view) and some of the images from the front, but rarely if ever images from the side or back. The model is well detailed and as far as I can tell is well overlapped from the 48 unique photos. I made masks for the model but it made little difference on the outcome. Any advice is appreciated.

The photo set I've been using with the masks included can be found here:

https://drive.google.com/drive/folders/1qQ4BCS071U8Au6nf2xtJjQER_lQOgnES?usp=sharing

[cam3d]

hi Dangerfro - Welcome to the forum.

Thanks for sharing those images - They are very low resolution (800*800px) and likely have JPG compression artifacts. Typical datasets will have 4920*3264px (16mpx) images - You can use less, but below 2560*1920px (5mpx) things can be a bit hit and miss, especially with compressed JPG files.

There is also a very steep angle between the top-down and side-on views, with not enough images in between to successfully orient the images.

The masks help minimally in this case as the background is basically featureless anyway.

I recommend taking pictures yourself so you have control over the setup, can take more photos, have images with a reasonable image resolution.

[Dangerfro]

I see, well regardless I'd like to try and make this image set work. I've had success with image sets similar in size and compression to this before and I'm not quite sure what's making this set such a bother.

[cam3d]

Dangerfro happy to have a look at a working set to compare to the non-working set and give you feedback - most likely it's down the surface features of the subject itself. Homogenous surfaces are inherently harder to orient cameras to using SFM processing, whereas a subject with lots of contrasting points of detail will work better. With such small datasets such as these, this becomes critically important, whereas with higher resolution images, you have more forgiveness with more data to reference from.