RetinaFace: Single-stage Dense Face Localisation, implemented in TensorFlow 2.0 | Hacker News Reader

RetinaFace: Single-stage Dense Face Localisation, implemented in TensorFlow 2.0

github.com

62 pointsstan_btd23 comments

Model | Easy | Medium | Hard

Mxnet | 96.5 | 95.6 | 90.4

Ours | 95.6 | 94.6 | 88.5

My professors would be so mad if I submitted data like this. I already can hear "What are the units? Seconds? So yours are by one second better? Error margin? Percent? So yours is worse?"

I get that this is a very specific information for a specific audience. People who stumble on this repo should know what is that.

However we can all be better at presenting our data.

as stated in the repo, its "mAP result values"

https://medium.com/@jonathan_hui/map-mean-average-precision-...

Don't get me wrong. I totally get it. But my professor is saying:

"So this is precision? 0 to 1? 1 being totally accurate? And you got 96.5? I assume it is percentage?"

It says "mAP result values on the WIDERFACE validation dataset:"

If your professor is working in on object detection they know what mAP is - all the major datasets use it as their standard evaluation criteria.

So is it mAP of 0.956 or 95.6? Why not 956?

Not to be rude, but if you can't work that out you shouldn't be working on this. I'm sure the author would take a pull request, but but it really sounds like you are nitpicking.

Figures like precision and recall are often expressed either as 66% or 0.66. Confusion really isn't that big a problem.

Not rude at all because I am totally nitpicking.

If you would write mAP: 0.9 or mAP (%): 90 or similar my "professor" would be OK with that.

I guess general rule in science (or maybe engineering) is to define units. I have no doubt that e.g. when you are building rockets (or maybe model airplanes) you would know that e.g. velocity is in m/s. Until somebody thinks in some other way and things blow up.

I just wanted to make some fun comment about missing units and that is it. I have no doubt that anybody browsing that repo would not be confused about mAP being 95.6.

It says mAP right above the table. If you know what mAP is then you know the units.

Professors will rail against imprecision to undergrads and young graduate students but they use it all the time in real life.

I can already hear my professors saying "stop commenting on style, what do you have to say about substance?"

The algorithm have been already shipped within the release of the PixLab Rest APIs 1.9.72: https://blog.pixlab.io/2020/08/pixlab-api-1972-released

Note however that Retina does not support real-time performance on the CPU especially on IoT devices and web browsers (WebAssembly). That's why we opted for a standard cascade approach for our WebAssembly port: https://sod.pixlab.io/articles/porting-c-face-detector-webas...

I couldn’t find any license text in the repo.

Would you consider using an open source license like for example the ISC license?

https://choosealicense.com/licenses/isc/

Of course, done !

Great, thank you :)

Not an expert, why is it called state of the art if its accuracy is worse than mxnet, whatever that is

The algorithm is state of the art. There are several implementations of the algorithm, the original is with the mxnet framework, and my implementation with tensorflow framework has a slightly lower accuracy.

Thanks! What makes it state of the art and why is lower accuracy acceptable?

The paper that presents this algorithm has the best known accuracy on the widerface dataset, which is why it is called state of the art. The authors of the paper published an implementation of this algorithm based on mxnet, but a lot people and companies use tensorflow instead of mxnet in their work, so just using the mxnet implemenation is not an option. Thats why I converted it to TF, with a slight decrease in accuracy on widerface. Then what is "acceptable" depends on your judgement, but the widerface dataset is extremely challenging, with many pictures having hundreds of small faces. In some of these pictures my implementation will miss a few faces, or find them but with lower probability. Overall for the vast majority of face detection applications, the two implementations will yield extremely similar results. I havent yet found a picture with a few dozen faces where the TF implementation performed not as well as the original one !

Got it. I understand now. Don't be offended at my use of "acceptable", it was used in ignorance.

no offense taken at all !

Any thoughts on distilling it or pruning to make it smaller ?

would you be interested in a smaller model file size or in a model that runs faster ?

For model file size I have a few ideas of things that could be tried easily.

For a faster model, this model is based on resnet50 architecture. The original authors also published a lighter mobileNet architecture. I havent converted it to tensorflow but I could try to take a look

> What makes it state of the art

Generally it means something roughly like "the best known approach for this specific problem". Often it means "the best known approach for this specific dataset" (eg "SOTA on ImageNet").

> why is lower accuracy acceptable?

Lower accuracy is worse, but these numbers look close enough that it's probably acceptable for most people.

There are plenty of environments where TF is preferable to MXNet (eg, you have TPUs/want to use TFLite on mobile/want to slice the model weights up and use it for your own custom TF model).

It's probably lower accuracy because it wasn't trained as long. Those extra couple of points could take days (or more) of training.

Awesome, thanks for the explanation!

I feel much safer now.

Reply on news.ycombinator.com