muy interesante, para artistas 2d como yo que deseamos dar el salto al 3d esta herramienta podría facilitarnos mucho la vida, estaré pendiente de su desarrollo
Brooooo What to do 🙁 Am getting the same error even tested with different pictures with simple backgrounds 🙁 Even after editing in photoshop with simple T poses. net = PoseEstimationWithMobileNet() checkpoint = torch.load('checkpoint_iter_370000.pth', map_location='cpu') load_state(net, checkpoint) get_rect(net.cuda(), [image_path], 512)
The Neural Network trained itself to estimate/guess the missing 3D aspects of this otherwise 2D image after it's "seen" thousands of images of similar height, gender, clothing, etc from all angles. You can see that it didn't estimate the jacket at the back accurately... Huge well-curated datasets have been put together for this purpose through the courtesy of social media and our mindless cooperation; data is truly the world's most valuable resource. This technology is still in its infancy but it's coming along at an astonishing pace. Edit: I just realized that *Star Man* gave pretty much the same explanation right before me.
Google. Google cams. Street cams. Instagram. Tik-Tok. FB. All free, public, given willingly and STUPIDLY. I figured it out on the first branch since MySpace, then Facebook... it was easy to see if you are aware, use critical thinking, know what Government actually is and who and what is controlling the world. You have to use your brain.
@@V.Z.69 Yeah too bad the world isn't that interesting, the AI just approximates what the back of each character it digitizes into a 3D Model should look like and arguably stlll does a pretty bad job at it for now.
I would be more impressed if all the surfaces were respected all the times, and not fused together when in contact. For example the fingers should remain separated even when the hand is closed, or the hand is leaning against the body. It can guess that becouse it saw the hand open sometimes.
I guess it's just a matter of time. I believe that making it temporally cohesive is a more important step. This paper is amazing and already unbelievably useful as is. With temporal cohesion it could get even more useful for several purposes. For example, someone would be able to use its output as an input to another neural net that could build a rigged and remeshed model with the animations.
how to use png sequence or any video?
This is very interesting indeed.
muy interesante, para artistas 2d como yo que deseamos dar el salto al 3d esta herramienta podría facilitarnos mucho la vida, estaré pendiente de su desarrollo
Brooooo What to do 🙁 Am getting the same error even tested with different pictures with simple backgrounds 🙁 Even after editing in photoshop with simple T poses.
net = PoseEstimationWithMobileNet()
checkpoint = torch.load('checkpoint_iter_370000.pth', map_location='cpu')
load_state(net, checkpoint)
get_rect(net.cuda(), [image_path], 512)
how the f does it know whats in the back
It's just a guess. As you can see it's pretty inaccurate and it will be improved in future versions.
The Neural Network trained itself to estimate/guess the missing 3D aspects of this otherwise 2D image after it's "seen" thousands of images of similar height, gender, clothing, etc from all angles. You can see that it didn't estimate the jacket at the back accurately... Huge well-curated datasets have been put together for this purpose through the courtesy of social media and our mindless cooperation; data is truly the world's most valuable resource. This technology is still in its infancy but it's coming along at an astonishing pace.
Edit: I just realized that *Star Man* gave pretty much the same explanation right before me.
Google. Google cams. Street cams. Instagram. Tik-Tok. FB. All free, public, given willingly and STUPIDLY. I figured it out on the first branch since MySpace, then Facebook... it was easy to see if you are aware, use critical thinking, know what Government actually is and who and what is controlling the world. You have to use your brain.
@@V.Z.69 Yeah too bad the world isn't that interesting, the AI just approximates what the back of each character it digitizes into a 3D Model should look like and arguably stlll does a pretty bad job at it for now.
Amazing 😍😍
Even if its not perfect, I believe an artist can do the final touches instead of starting from scratch. This will save 70% of artist time
how to use video input????????? aaaaaahhhhh!!!! love it
This is incredible
This is awesome !
Please tell me how to do it
It’s very interesting research, but all those flickering polygons, it’s far from production use I guess
Hi.How can i buy this programm?
Feed the artist behind.
I would be more impressed if all the surfaces were respected all the times, and not fused together when in contact. For example the fingers should remain separated even when the hand is closed, or the hand is leaning against the body. It can guess that becouse it saw the hand open sometimes.
I'm sure it will get there soon. Still an interesting start.
Well, that will come in the future, if I'm not mistaken it is trainable, so it'll eventually get there.
I guess it's just a matter of time. I believe that making it temporally cohesive is a more important step. This paper is amazing and already unbelievably useful as is. With temporal cohesion it could get even more useful for several purposes. For example, someone would be able to use its output as an input to another neural net that could build a rigged and remeshed model with the animations.