Name: DeepFace AI
Uploaded: 2025-01-22T12:30:51.809Z
Duration: 2 min 33 s
Description: A Python framework for facial recognition and attribute analysis, supporting multiple state-of-the-art models.

deepface-ai

A Python framework for facial recognition and attribute analysis, supporting multiple state-of-the-art models.

DeepFace AI

Hi folks. This is Shafik. Today we are going to focus on the recent research we published about deepface library. In this research paper, we basically focused on the best configuration set of the PICE including facial recognition model, face detector, distance metric, and alignment B. 

The research paper we are going to talk about is available under the repositories R file. You can access it with this BCH or under the citation section. You're going to find it beex. This article is published in Gazi University's The International Journal of Information Technologies and also this work is published under the supervision of Associated Professor Aler Espar. The content of the article is available publicly. 

Of course, you don't have to read all this article, but if you skip.

The tables page 102 you are going to see six different tables. Columns are representing facial recognition models and rows are representing face detectors. Also, another two dimension values which are distance petric and alignment mode are shown as separate tables. 

As you may know, the p is working like a automatic car or flight with autopilot. What I mean, if you call at verify function, you just need to send the exact pets of image pairs here image 1.

Geg and image 2.

Geg are the EXA pads and this is going to return the result you don't have to set anything else but of course you can tune it like a manual car. 

And let's focus on the source code of the Bas for example if you don't set anything to model name then the model is going to be W face but you can set model as face net face net 512 open face deep face deep ID D lip Arc face s face and Ghost Face net. 

Also similarly if you don't set anything to detector back end then open CV face detector is going to be used but if you want to tune you can set it to retina face mtcnn SST DP media pipe YOLO Center face or you can skip the face detection step. 

What's more for distance metric the default values cosine similarity cosine distance but alternatively you can set it to Idan or L2 normalized Idan distance. 

And finally align mode is active by default but you can set it to false to disable it. 

So in this research we basically investigated the results of this configuration sets as I.

Mentioned distance metric is cosine distance. It's cosine distance and enabled alignment. This table is going to be run with the default configuration. Also, the default facial recognition model is VGG Face and the default face detector is Open Face. In other words, if you perform base.

Verify wet just ex image pads then you are going to get 87.

Point per accuracy score on labeled face in the wild data sets test data set. On the other hand, as mentioned in the rhym file, human beings accuracy score on same data set is 97.5. So if you run deepface with the default configuration, this is going to show slightly lower performance than human beings. 

But as you can see in this tables, some values are highlighted boldly. This means these configuration sets past the human level accuracy. The best score retrieved from the configuration sets is 97.4. We reach that score in the last two tables, and both are coming from face net 512 facial recognition module and retina-face face detector. 

So if you want to reach the best accuracy score in the face Library, you should use faet 51 to facial recognition model, retina-face face detector, cosine distance or idian distance as distance magic, and enabled alignment mode. This is 0.9 almost 1% higher than the human level accuracy. Car also another important outcome of this research is mentioned.

In the read me file of three poster please visit the facial recognition models section in the r m. Then you are going to see a table. These are the reached or measured best cost of each facial recognition model. 

As you can see just face Net 5 mon to pass the human level accuracy score. On the other hand face net regular face net is coming close to the human beings. The lip vgg face and Arc face have acceptable scores. They are very close. 

Face net s face shows slightly lower performance than the acceptable scores. Finally open face deep face and deep ID models show poor performance. You should focus on the measured score column because these are measured from Deep Bas library. 

On the other hand original papers of these models uh may show different score values. For example deep face shows 97.3% accuracy score in the original paper but this is very lower than the score in the Deep face library. Because deep face library is using an open-source implementation of the Bas model.

Even though it's backbone model is shared publicly, it's weights are not published publicly. So we are adopting it open source implementation. That's why we cannot use the real model or V G face published its print weights. But we have different score than the original paper. This may be because of the adopted PR normalization technique or detector technique. This is mentioned in this paragraph. 

So you should focus on the measured score to see the reach scores of facial recognition models from Deep Face. As mentioned here, you should see the benchmarks page. Because, uh, just after we published this paper, we added a couple of new facial recognition modules and face detectors. For example, Ghost Face net is not available in the columns here. Also, YOLO and Unet or Center Face or Fast Empty CNN are not available in the rows. 

So you can see the whole table, extended table of the research paper in this page. Here you are going to find the nearly added Ghost Face net, YOLO, Unet, Center Face models and.

Still the best score is 97.4. And you can see the row curves of different facial recognition models for their best configuration set. And as you can see, FaceNet 512 is the best model in the Deep Faces portfolio. 

Finally, you should visit the evaluate results notebook. Uh, there are important outcomes in this notebook file. Basically, here we are showing the accuracy scores. But here, for example, alignment impact. If you would not adopt the alignment mode, you may remember that at default value is true. But if you set this to false, then you are going to lose. You are not going to have these improvements. 

In ArcPhase, alignment mode improves the score almost 6%. Similarly, what the detection impact over the accuracy? The best score comes from Arace again, and it's up to 40%. It's very important. 

So, in other words, we can understand that detection and alignment are very important stages of facial recognition pipelines. Also, you can see that if you skip the detection step, for example, with FAET.

Module you are going to get almost 60% score. But if you started to use the detector, then the score jumps to 90%. So in this video, we focus on the recent research paper of the face Library titled A Benchmark of facial recognition pipelines and Co usable performances of modules. 

In this research, we are basically focusing on the impacts of configuration sets in the face Library. Finally, source code of the study is available under the benchmarks folder. If you run perform experiments first and evolute results second, then you are going to get the same results. 

I hope this research will contribute to the academic Feld. Thank you all for watching and see you next time.

Note, that there may be slight differences in text and video. Watch video for the original content.

DeepFace AI Video