human.tf.memory().numTensors after human.load() doesn't give zero about human HOT 5 CLOSED

Models are also using tensors! So only time you'll get zero is if all models are disabled.
The goal is thar number does not grow over time as that would indicate a memory leak.

Its also normal that it varies during usage (temp tensors, etc.), but it should always return to same value once it's done.

from human.

Thanks for the clarification, do you have any tips or advise for optimizations to be speed up detection and make it less resource intensive if I am using human in the context of a chrome extension?

from human.

i'm not sure what to add, you're already using webgl, number of models is minimal and cache sensitivity is set very low.
only major impact would be to reduce maxDetected as faceres that you have enabled runs for each of detected faces.

from human.

what about the image size? currently I am not using the filters to reduce image size because I am not sure how to maintain aspect ratio, instead I am using my own function where basically I wanna reduce the image size if the width is greater than a max width or the height is greater than a max height otherwise no reduction is needed, but using filters I can provide either the max width or the max height because if I provide both to the filters then it will just resize to them and will distort the aspect ratio, is my understanding right?
Also, what's the smallest you'd recommend resizing the image to before you notice a decrease in the accuracy? given that I am using blazeface back model

from human.

resolutions on which each individual models have been trained on are in wiki - specifically faceres is trained on 224px, so your desired input resolution should be such that cropped face is ~224px - more that that and its pointless, less than that and you loose precision. what does that mean for overal resolution? well, depends if faces are near or far from the camera.

but yes, higher resolution does come with performance impact - and not only inside human, due to offscreen usage there is extra copy to transfer image context to human and that is not as fast as i'd like it to be in chrome - but it cannot be avoided.

regarding filters and aspect ratio - its handled automatically if you provide just width or just height.

from human.