Tweet
Hi there,
For my graduation research I am researching low-budget methods for reconstructing/scanning a human face in 3D. The aim is to find solutions that are aprox. 1000,- eur.
I have come to the conlcusion that photogrammetry might be a relative cheap option, especially when using a smartphonecamera.
To reconstruct human faces, I have seen, using lots of photocamera's in a rigid setup, is probably the best way. I have gone another route, though, because this is still an expensive solution (even when using lots of Pi-cameras). I was inspired by using video, especially now 4K-video is becoming more mainstream. So I have made an construction where 2 camera's are mounted on a tripod, which in turn is mounted on a dolley with a rail. So the camera's can be pushed around 180 degrees in one swift movement. I have achieved acquisition time of aprox. 3 seconds in this setup.
Now, I have made some 3D-models with 3DF, while also making a 3D-model with a 3DmD scanner, which I want to use as Ground Truth. (Ofcourse the 3DmD scanner has it's own inaccuracy, but this one of the best methods I can use, right now.) And I have made 3D-models of the faces using Sony Xperia 3D-creator app.
I want to compare these models with the "Ground Truth" model so I could approximate the (in)accuracy of my 3D-models, made in my setup.
I have read that using ICP is probably the best way, but using ICP on 3D-models of faces is a very hard way, I found out, because of the variability of the faces.
Cutting the 3D-models to the same size and registrate them precisely, is an very hard task. Also, I can't get my 3D-models to the same scale, because placing rulers, or using geolocation is almost impossible.
So using ICP to get an actual usable number, is not possible. (Please prove me wrong...)
So the best way I can think of righ now, is placing markers on the face with a pen. After 3D-reonstructing, placing landmarks in the 3D-model, on top of the texture, and then measure the distances between these landmarks in the different models and compare them to the distances of the ground-truth model. For reliability sake, letting multiple people place the landmarks in the 3D-models.
I know this isn't the best method as well, because with this method I would omit the rest of the 3D-model, between the points. So it would'nt be the best representations for the accuracy of the whole model, but only for the landmarks themselves.
So, my question to you is, how would you aproach this problem, in getting the actual (in)accuracy of your 3D-model? Maybe somoeone can give me a different perspective.
Best regards,
Arnand
P.S. More questions about the setup are welcome
.
For my graduation research I am researching low-budget methods for reconstructing/scanning a human face in 3D. The aim is to find solutions that are aprox. 1000,- eur.
I have come to the conlcusion that photogrammetry might be a relative cheap option, especially when using a smartphonecamera.
To reconstruct human faces, I have seen, using lots of photocamera's in a rigid setup, is probably the best way. I have gone another route, though, because this is still an expensive solution (even when using lots of Pi-cameras). I was inspired by using video, especially now 4K-video is becoming more mainstream. So I have made an construction where 2 camera's are mounted on a tripod, which in turn is mounted on a dolley with a rail. So the camera's can be pushed around 180 degrees in one swift movement. I have achieved acquisition time of aprox. 3 seconds in this setup.
Now, I have made some 3D-models with 3DF, while also making a 3D-model with a 3DmD scanner, which I want to use as Ground Truth. (Ofcourse the 3DmD scanner has it's own inaccuracy, but this one of the best methods I can use, right now.) And I have made 3D-models of the faces using Sony Xperia 3D-creator app.
I want to compare these models with the "Ground Truth" model so I could approximate the (in)accuracy of my 3D-models, made in my setup.
I have read that using ICP is probably the best way, but using ICP on 3D-models of faces is a very hard way, I found out, because of the variability of the faces.
Cutting the 3D-models to the same size and registrate them precisely, is an very hard task. Also, I can't get my 3D-models to the same scale, because placing rulers, or using geolocation is almost impossible.
So using ICP to get an actual usable number, is not possible. (Please prove me wrong...)
So the best way I can think of righ now, is placing markers on the face with a pen. After 3D-reonstructing, placing landmarks in the 3D-model, on top of the texture, and then measure the distances between these landmarks in the different models and compare them to the distances of the ground-truth model. For reliability sake, letting multiple people place the landmarks in the 3D-models.
I know this isn't the best method as well, because with this method I would omit the rest of the 3D-model, between the points. So it would'nt be the best representations for the accuracy of the whole model, but only for the landmarks themselves.
So, my question to you is, how would you aproach this problem, in getting the actual (in)accuracy of your 3D-model? Maybe somoeone can give me a different perspective.
Best regards,
Arnand
P.S. More questions about the setup are welcome
.
Comment