Comment by danielmorozoff | Hacker News Reader

Comment by danielmorozoff | Hacker News Reader

It really depends what you are attempting to accomplish, and what you wish to detect in the images.

As you mentioned faces:

Are you looking for face detection or recognition? Face detection has been robustly solved before the advent of DL with HAARs/ face models. Now being pushed a bit further with DL.

(http://docs.opencv.org/master/d7/d8b/tutorial_py_face_detect...)

Current cutting edge face recognition systems rely on DL, and the top performing models are one out of Russia (NTechLAB, facenx_large) and one from Google (FaceNet v8). These were the top two performers in the MegaFace challenge - identification with 1M distractors. Truly remarkable results. http://megaface.cs.washington.edu/results/

As with most DL systems you will need a massive corpus of labeled faces (aka, google or vkontakte - which the NTechLab group used)

Note from personal experience: Haar cascades only work really well on frontal faces in high quality and good light. It sounds like OP has these kinds of photos so it will work well, but if you want to detect faces in any other kind of image/video, you'll need something more powerful. I still haven't found anything that works well.

You may want to check out the face detection approach described in this paper: Joint Cascade Face Detection and Alignment. ECCV 2014(http://www.jiansun.org/papers/ECCV14_JointCascade.pdf)

I'd also encourage you to try out the Face API from Microsoft (full disclosure, I work on it). One of the focuses is on improving detection when challenging lighting and occlusions are present: https://www.microsoft.com/cognitive-services/en-us/face-api

Very interesting paper. Thank you for sharing it.

I had a question after reading it. The parameter rho, that is used to either classify/regress depending on the t cascade stage{1..T}, is said to be set empirically. This, I would assume, could change across data sets, how were you able to decide when classification switched to regression in your tree and how was this adapted when testing on other data?

Also forgot to attach the paper I was referring to: http://arxiv.org/pdf/1502.02766v3.pdf

Reply on news.ycombinator.com