Performance analysis of DLib’s CNN face detection

For the purpose of this application, we did a small analysis of how exactly we need to scale/adapt images before feeding them to DLib's CNN face detection algorithm. You can find performance analysis here:

https://github.com/stalker314314/dlib-cnn-face-detection-analysis

Results for the project.

• The use of ram memory is limited to the PHP configuration.
• If it is not configured, try to use half of the system's memory.
• Regardless of these values, no more than 4Gb of ram will be used. Using more memory, implies more processing time, and does not improve the results.
• Not allow any execution if there is less than 1GB of RAM available to work.

Keep in mind that these limits are purely speculative. That is, a file size is selected for the analysis, which in practice when analyzed, consumes at most this amount of RAM. This result, which in the practice of our developers works, may vary for you.

On the other hand, if you set a high memory limit, this process will try to use it, and is independent of the current system load. If your system load is high, you may not have enough memory to continue the analysis, and this process can be stopped or killed due to OOM by the Kernel.