search

uricamic / clandmark

Open Source Landmarking Library
http://cmp.felk.cvut.cz/~uricamic/clandmark
GNU General Public License v3.0
199 stars 111 forks source link

Why the detected landmarks in the example "static_input" are not correct? #57

Closed wguo68 closed 8 years ago

wguo68 commented 8 years ago

bad

And is there any trained model to detect more landmarks as in your website?

uricamic commented 8 years ago

Hi,

the reason is that static_input (unlike video_input) uses new mechanism of features computation (CFeaturePool), which makes the flandmark_model.xml incompatible.

Sorry for the inconvenience, I will upload the new flandmark_model.xml compatible with CFeaturePool soon.

uricamic commented 8 years ago

I forgot to add, that you can of course still use flandmark_model.xml, but you have to remove CFeaturePool from the code (look at video_input.cpp example, which is using the old style).

Or you can use the new style, but with the other models available for download on the project webpage (like CDPM.xml, FDPM.xml, jointmv*, etc.).

wguo68 commented 8 years ago

Where is the other trained models for download ? Is there 68 landmark trained model?

There is an compile error for the code line "!face_cascade.load( argv[1]+face_cascade_name ) " .I modified this line to "!face_cascade.load( argv[1]) "

uricamic commented 8 years ago

Hi @wguo68,

the learned models are here.

CDPM.xml and FDPM.xml are detecting 68 landmarks and they are supposed to be use together, forming the C2F-DPM, check the code snippets for this kind of detector.

wguo68 commented 8 years ago

It's Okay. Great! good

For large image，it may not be correct enough. How to improve the result for large image?

uricamic commented 8 years ago

Hi @wguo68,

sorry for late answer, I didn't get notification for a new message in this thread.

For larger image, there are several options what to do. If you would like to use only CLandmark, you could train a new model for a higher resolution images, however this would take a lot of time and resources. Other possible solution is to use CLandmark for initialization of some other technique which is more suitable for high resolution images (with large faces), like the Active Appearance Models (AAM) and related methods. These are known to have very good results for such images, however they usually suffer when the initialization is poor. The combination of CLandmark and AAM could provide fast, but still highly accurate detection.