4 ms·

Looks like more evidence for Atwood's Law: Any application that can be written in JavaScript, will eventually be written in JavaScript. I have a couple of ques

by bfrs 15y ago

Looks like more evidence for Atwood's Law: Any application that can be written in JavaScript, will eventually be written in JavaScript. I have a couple of questions: 1. How does Liu Liu's algorithm (which I think he calls "deformable parts model") compare with Viola-Jones [1]? Any idea what the error stats (false positives & negatives) are like? In my experience [2] Viola-Jones has very low error rates. Regarding performance, I think it'll be a while before anything can come close to Viola-Jones, but this algo seems good enough. 2. Ok, this might be a stupid question, but I'll ask anyway. Does this plugin mean that face detection can now be completely done on the client side? I use face detection in my web app [2] for passport photo formatting. I'm wondering if instead of uploading the photo (~2-4MB) to the server and processing it there with OpenCV, I can download the plugin to the browser(~0.5MB?) and get everything done right there. [1] http://opencv.willowgarage.com/documentation/python/cascade_classification.html http://opencv.willowgarage.com/documentation/python/cascade_... [2] http://freepassportphoto.dyndns.org/ http://freepassportphoto.dyndns.org/ The error rates info is from the webserver's error-log. Its very rare that I see a "No face found" error (false negatives) or get complaints that the formatted photo is all messed up (false positives).

SammoJ 15y ago
The deformable part based model is pretty much /the/ state-of-the-art object detection model. There is absolutely no way it should be attributed to "Liu Liu". It's creators are Felzenszwalb et al.[1]. The performance of DPMs far surpasses any other basic model such as a Viola Jones syle Haar cascade. Also you should find with DPMs that you are able to detect faces with far higher degrees of deformation - i.e. 3/4 profile faces, rotated faces etc. DPMs are far far more robust than haar cascades but with greater computational expense. Check out the papers at [1]. [1] http://www.cs.brown.edu/~pff/latent/ http://www.cs.brown.edu/~pff/latent/
- bfrs 15y ago
  I was using "Liu Liu's algorithm" in a colloquial way. Should have been more specific and used, "the algorithm which Liu Liu implemented". As far as robustness is concerned I guess a lot depends on the quality of the training. The haar-caascade used in openCV is very well trained and generates few false +/-'ves. I wanted to know how this particular implementation of DPM performed.
etherealG 15y ago
Liu Liu's code is a javascript port from openCV I think. As far as I know it is an implementation of Voila-Jones.

by bfrs 15y ago

SammoJ 15y ago

The deformable part based model is pretty much /the/ state-of-the-art object detection model. There is absolutely no way it should be attributed to "Liu Liu". It's creators are Felzenszwalb et al.[1]. The performance of DPMs far surpasses any other basic model such as a Viola Jones syle Haar cascade. Also you should find with DPMs that you are able to detect faces with far higher degrees of deformation - i.e. 3/4 profile faces, rotated faces etc. DPMs are far far more robust than haar cascades but with greater computational expense. Check out the papers at [1]. [1] http://www.cs.brown.edu/~pff/latent/ http://www.cs.brown.edu/~pff/latent/

bfrs 15y ago
I was using "Liu Liu's algorithm" in a colloquial way. Should have been more specific and used, "the algorithm which Liu Liu implemented". As far as robustness is concerned I guess a lot depends on the quality of the training. The haar-caascade used in openCV is very well trained and generates few false +/-'ves. I wanted to know how this particular implementation of DPM performed.

etherealG 15y ago

Liu Liu's code is a javascript port from openCV I think. As far as I know it is an implementation of Voila-Jones.