Can't reconstruct turntable dataset

[Serge]

Hi! I have circular sequence of PNGs with alpha channel working as mask. But any settings I've tried end up with too many discarded images or their wrong orientation. Some old thread here suggested to keep keyfeature detection level and matching depth low along with high matching accuracy, if surfaces are undertextured. Done! But neither sequentional nor circular modes orient cameras well. Out of ideas with this set...

Zephyr 3.7 free version

Link to the dataset (30mb), 36 images

[Andrea Alessi]

Hi Serge,

i had a quick look at your dataset and yes, unfortunately the subject is not very textured and moreover images are very small (1300x1300). Do you have the original images by chance? if the turntable itself that moved with the object was texture enough, ther emay be some way to salvage the dataset

[Serge]

This quality is only one I have, wish it was better. I was thinking about manually placing cameras in circle at every 10 degrees - is it possible, and reconstruct from given point. Also I did a quick test with replacing camera images in another project (where cameras are already lined in circle). But I couldn't re-generate sparse cloud with new images thus point dense cloud become very rough

[Andrea Alessi]

Hi Serge,

apologies for the late reply. Vbulletin had mistakenly tagged your post as spam and i had to manually approve it.

Yes, replacing images from a pre-calibrated setup can work but obviously you need a fixed rig. However, if you're planning to scan those type of subjects, i would recommend to improve the photo quality as much as you can.