FIXED: FaceTrack CHOP - "file could not be found" error

I really really appreciate it. Thank you. With 2021.12569, FaceTrack CHOP works with RTX3090, win10, NVIDIA driver version 461.92.
(and it did not work with NVIDIA driver version 457.51)

SUCCESS! works great on my Win10, RTX3070, Driver version 461.92

1 Like

For me it works also now, only odd thing is it drops out when I disable the viewer active on the Face Track CHOP, but I really like this new feature :slight_smile:

Glad its working.

I don’t think the Face Track SOP is forcing the CHOP to cook at the moment, so if you’re not using any of the CHOP’s channels (or have the viewer off) it likely won’t update. I’ll get that fixed in the next update.

1 Like

@robmc Thanks a lot, it works now. Only begining to test it.

Downloaded and installed this version (i’ve 3090rtx)… the chop works, but when i connect the sop, face track chop give me this error:

P.s. i’ve checked all the buttons but nothing change.

Unfortunately, that error message is coming from the nvidia library and doesn’t really say much about what the actual problem is. That is the point where it is trying to load the TensorRT (TRT) ai model.

Possibly try an absolute path to your mesh file. Since nvidia is loading that file internally, sometimes the base path it is using is not the same as TouchDesigners.

You mentioned installing the cuda and ar SDKs - have you updated your graphics driver as well?

1 Like

all is updated… i’ll do other test in next days… thanks anyway :slight_smile:

ok. I’d definitely recommend trying it with an absolute path to the mesh file i.e. c:\program files\derivative\ …\meshfile.nvf I get the same error message you were seeing when trying with a relative path

Let me know if you’re still having issues.

1 Like

it’s working great on my machine using RTX2070, great stuff to experiment with - so far we had to pay a hefty license fee for a facetracking SDK with very similar features, so I’m curious to see how this version holds up.

@robmc I tried to filter the Facetrack CHOP data to make the SOP movements smoother - but I noticed I’m only allowed to reference the Facetrack CHOP in the Facetrack SOP. I’d like to put some Filter CHOPs in between (different filters on different face parts) and then reference the Null CHOP at the end as “Facetrack CHOP”.
Perhaps this option can be added to any future releases.

@nettoyeur If I understand, you’re looking to be able to alter the vertex data of the mesh in additional CHOPs before it becomes a mesh in the FaceTrack SOP? I guess it would technically be possible, but it’s also probably a little awkward in our SOP/CHOP workflow since I’m not sure how you would pass the triangulation data along and you’d end up with channels of different lengths (or 60,000+ channels with one sample if each vertex was separate).

At the moment, I think the FaceTrack SOP would also still have to reference the original FaceTrack CHOP as well as the Null so it would know how to build the mesh from the vertex positions (or we keep a separate CHOP for the triangle indices).

I will keep this in mind going forward. Nvidia’s FaceTrack system is still in beta so I definitely expect there will be further updates to this system.

@robmc hmm perhaps that’s not what I mean - I admit I’m not sure on what data the CHOP is sending under the hood, but I’m only looking to filter the transformation/rotation data in the visible CHOP channels- for instance the face:rx, face:ry and face:rz value, on the most basic level.
I assumed the other 140 single-sample CHOP channels (when using 68 landmarks) were only transformation/rotation data for bones/parts of the mesh like eyelid, left corner mouth etc. If I could filter those to, for instance: add some lag to the lips transform/rotate channels I could make the lips tremble less, with a Math CHOP set to multiply 1.5 exaggerate the movement influence on left eye, etc

Ah, thanks for the clarification. When you mentioned wanting to use different filters on different parts of the face, I assumed you were looking to manipulate the vertex data of the face because that’s generally all I have access to. Internally, I give the AI model an image frame and as output I get the bounding boxes, landmark points and also a list of vertices and indices that I use to build the mesh. The vertices are already pre-deformed by the AI so there isn’t currently any opportunity to further adjust them based on the landmarks before they become a mesh.

The only exception is the transform that aligns the entire mesh with the face in the video and that can be disabled and done manually by turning off the ‘Pre-Transform Mesh’ option on the SOP.

That being said, while I’m currently running the AI model in a single step, the SDK does support potentially breaking it up unto 3 separate stages. In this idea you’d actually chain together multiple FaceTrack CHOPs which would be a little slower but would allow you add filters or other logic at each step.

Specifically, you could have one FaceTrack CHOP that produced the landmark data, which you could then process via other CHOPs and send as input into another FaceTrack CHOP that would only run the last stage of the AI that produces the mesh. Your FaceTrack SOP would then be connected to that last FaceTrack CHOP.

I was originally trying to keep things fairly simple with the CHOP because it was a bit of an experimental side project, but I like the flexibility this could allow. I’ll definitely keep it in mind when I get a chance to work on this again.

1 Like

Hello, you and I encountered the same problem “The file could not be opened”, how did you solve it in the end?

Just updating to the new version. But I doesnt tested it further.

1 Like

Any updates with this?

I’m having the same issue, the CHOP works but when connecting the Face Track SOP it shows the “Error while loading the TRT model”, I have CUDA and AR SDK installed (Notch Body Tracking is working for example).

Have you tried the suggestion of using an absolute path to the facemesh model? e.g. c:\users\myname\meshfile.nvf?

Yeah, but it still gives me the same TRT model issue