Original article can be found here (source): Deep Learning on Medium
BlazeFace: Sub-millisecond Neural Face Detection on Mobile GPUs
Now we able to achieve fast speed face detection in mobile devices → very powerful and good. (they are able to capture the instant (they also developed another method than non-maxima suppression).
In mobile applications, → we start with face detection. (Hence this has to be very fastly done). (since we only want to put the other methods → such as face localization → in only the frames where there is a face).
A compact feature extractor was made as well as postprocessing.
Additionally, they developed different models for front as well as backend camera → which is quite amazing.
Since there is very limited space → the convolution operations must be done that cover larger image locations. (hence they decided to use a depthwise separable convolution operation).
A new block is made → this Blaze Blocks are made to be much more efficient. (increasing the kernel size → in DW convolution are cheaper)
The connection was made to overcome the smaller channel size. (Double Blaze Blocks were the main building block). (another novel anchor scheme was made and implemented).
Even in post-processing, → they developed a new method for better and faster post-processing. (used 66k images → and used private data for validation)
Still an amazing approach!
When compared to the mobile network → it gave a faster yet accurate prediction. (but mobile network covers much more classes).
The Blazeface covers a lot of areas where → the facial prediction was made. (Additionally, they have some trick in their training scheme).