About Flandmark precision

[guo253]

Hi, When i used flandmark_simple_example.m with flandmark_model.xml in data folder to detect facial points of a pic downloaded from internet, I found that the precision was not very well,especially the face was not frontal. The result was not well using joint_mv_model and independent_mv_model also. Is there any constraints about pic sizes? Or ,do you have more precisly model? My pic sizes is 114*136. Thank you.

[uricamic]

Hi @guo253,

could you please provide some output images?

The resolution is usually not a problem, there is just a systematic error introduced by re-scaling the input image to a working image (normalized frame).

However, there could be other problem with the compatibility of the models and the detection method used. We have made some speed ups for features computation, which results in a different scheme of computations and therefore the model has to be learned in the same way. It will be described on the webpages soon.

All multi-view models are using the new optimized way of feature computation, and therefore they require featuresPool class as well.

flandmark_model.xml is old model, which does not use this optimization. It was learned on near-frontal images only, so it works meaningfully only for faces with yaw in range around (-30, 30).

[scdeng]

Hi @guo253 and @uricamic I had the same precision problem, I ran the example with command "./static_input ./ ./flandmark_model.xml ./face.jpg face_result.jpg", the result seems not accurate as flandmark did. is there something wrong with the model in clandmark example?

[uricamic]

Hi @scdeng,

yep, the flandmark_model.xml which is in the package now is not compatible with that example. I will prepare new model identical to the uricamic/flandmark but compatible with the new scheme of features computation.

[futurely]

The test results on the 300-W IBUG dataset using the provided jointly learned models for multi-view facial landmark detection are completely inaccurate. Two of the best ones are shown below. Could you add some examples to train the models?

[futurely]

The individually learned multi-view models are even worse.

[uricamic]

Hi @futurely,

I think the problem is that you use incompatible models for your code. Please send me the images without detections so I can check it out.

[futurely]

The test images are the IBUG database downloaded from here.

[uricamic]

I meant just these two images you posted here. Or at least their file names.

[uricamic]

Hi @futurely,

here are the results when the correct code is used:

When in-plane rotation is returned by the face detector:

For the first image my face detector missed the face.

I will soon update both webpages and the repository with examples covering the multi-view scenario (as well as some new stuff).

Sorry for the inconvenience caused by the compatibility issues between old and new models. I will try to clarify this on the webpages as well.

[karlhugle]

Hi @uricamic the same issue, can't run as you showed above, it is more like the old landmark version. What is fd con:56.613 mean? is it a build-in function?

[uricamic]

Hi @karlhugle,

the fd conf: is a confidence of our face detector, nothing in connection to clandmark itself. As for the precision issues, there was some update of the code snippets and also there is another thread here: https://github.com/uricamic/clandmark/issues/21

The update of the webpages will happen next week, I will also include more self-contained examples.

[karlhugle]

@uricamic Many thanks！