In this post we are discussing face detection as a separate step. Refer to the post on "Better face suggestions" for a discussion of face recognition improvements planned.
If we continue to improve the accuracy of the face detector module, then that saves you time in two ways;
Factors in face detection include; resolution, lighting, pose, blur, occlusion.
We are testing a new machine learning face detector that is more accurate (>20% improvement). The downside is that it is slower and takes up lots of memory... so we are trying to minimize the negative impact before we roll it into production.
Tips and tricks:
- our minimum face size is approx. 100 pixels in width
- generally it is less effort (keyboard clicks) to delete false positives then it is to uncover and add false negatives.
- one trick to note - by doubling the image size (simply by blowing it up) so that the smallest faces are greater than 60 pixels in width that can improve the results in most cases if faces are missed on small pixel sized images (roughly a min width of 800 pixels is something to shoot for). a good tool for that is XnConvert or XnResize - in the XnViewMP family.
- we have embedded certain default settings for face detection confidence and quality
- if you have high resolution images then you can go with much higher confidence settings to avoid false positives
- by decreasing the confidence score more faces are picked up along with more errors; that is ok for low resolution images from email attachments and old scans of hardcopy photos; but not so good for higher resolution images.
- if you would like more information on these low level settings we can share them with you and how to "tune" them. but we suggest that you only do this if you are technical and are confident editing configuration files.
Another tip is to enhance the quality of your images - if you are dealing with old scans. One tool available is available at MyHeritage - example enhancement steps below.