Looking for ways to explore internal libraries of Compreface

[SarimZH]

Hi there, im trying to use Compreface recognition for a project. However, the compreface recognition service returns false positives with a certain high similarity index (>=0.9) on South Asian faces. The project needs accurate recognition since it involves transactions. Similarly, it also creates log files which ends up consuming a lot of space.

What I am looking for is to access the faceprint code within the container which would somehow simplify and accelerate the recognition process.

[pospielov]

Hi, if you look at our architecture: https://github.com/exadel-inc/CompreFace/blob/master/docs/Architecture-and-scalability.md compreface-core is responsible for calculating faceprints compreface-api uses these faceprints to calculate euclidian distance. I am not sure that you can improve the accuracy without changing the model that calculates faceprints. The problem is that if during the model training it didn't see enough faces of a particular race, it will show low accuracy and I don't see how you can improve it without retraining the model. Also, have you tried the InsightFace custom build? It should show better accuracy on Asian race (Still not the best)

[SarimZH]

Hi, thank you for your response. In that case, is there a way that I can replace the model in compreface container with the Insight face model?

On Fri, Feb 18, 2022 at 8:02 PM Pospielov Serhii @.***> wrote:

Hi, if you look at our architecture:

https://github.com/exadel-inc/CompreFace/blob/master/docs/Architecture-and-scalability.md compreface-core is responsible for calculating faceprints compreface-api uses these faceprints to calculate euclidian distance. I am not sure that you can improve the accuracy without changing the model that calculates faceprints. The problem is that if during the model training it didn't see enough faces of a particular race, it will show low accuracy and I don't see how you can improve it without retraining the model. Also, have you tried the InsightFace https://github.com/exadel-inc/CompreFace/tree/master/custom-builds custom build? It should show better accuracy on Asian race (Still not the best)

— Reply to this email directly, view it on GitHub https://github.com/exadel-inc/CompreFace/issues/730#issuecomment-1044664277, or unsubscribe https://github.com/notifications/unsubscribe-auth/AXLJ3D22F5VYKIMCAQJEHXLU3ZNQPANCNFSM5OXSC6NQ . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

You are receiving this because you authored the thread.Message ID: @.***>

[pospielov]

The model is placed in slightly different places in each build, here is an example for exadel/compreface-core:0.6.1-mobilenet-gpu build: /app/ml/.models/insightface/calculator/arcface_mobilefacenet I think you easily will find the same path for other builds. The problem is that you can divide facial recognition into two phases - faceprint detection (this is what this mode does) and comparing faceprints to find face similarities. The second phase has its own coefficients, so the resulting similarity always will be between 0 and 1. So when you replace the model it will work, just prepare to see very strange similarities.

[SarimZH]

Hi there,

Thank you for guiding me about this. Could you also tell me if Mobilenet Custom built would work too for Asian faces?

On Fri, Feb 18, 2022 at 8:02 PM Pospielov Serhii @.***> wrote:

Hi, if you look at our architecture:

https://github.com/exadel-inc/CompreFace/blob/master/docs/Architecture-and-scalability.md compreface-core is responsible for calculating faceprints compreface-api uses these faceprints to calculate euclidian distance. I am not sure that you can improve the accuracy without changing the model that calculates faceprints. The problem is that if during the model training it didn't see enough faces of a particular race, it will show low accuracy and I don't see how you can improve it without retraining the model. Also, have you tried the InsightFace https://github.com/exadel-inc/CompreFace/tree/master/custom-builds custom build? It should show better accuracy on Asian race (Still not the best)

— Reply to this email directly, view it on GitHub https://github.com/exadel-inc/CompreFace/issues/730#issuecomment-1044664277, or unsubscribe https://github.com/notifications/unsubscribe-auth/AXLJ3D22F5VYKIMCAQJEHXLU3ZNQPANCNFSM5OXSC6NQ . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

You are receiving this because you authored the thread.Message ID: @.***>

[SarimZH]

I had accessed the compreface-core container and was able to look into the .py files of InsightFace and Facenet. However, I could not find the yml file to replace it with custom models available.

On Tue, Feb 22, 2022 at 11:14 AM Sarim Zuhair @.***> wrote:

Hi there,

Thank you for guiding me about this. Could you also tell me if Mobilenet Custom built would work too for Asian faces?

On Fri, Feb 18, 2022 at 8:02 PM Pospielov Serhii @.***> wrote:

Hi, if you look at our architecture:

https://github.com/exadel-inc/CompreFace/blob/master/docs/Architecture-and-scalability.md compreface-core is responsible for calculating faceprints compreface-api uses these faceprints to calculate euclidian distance. I am not sure that you can improve the accuracy without changing the model that calculates faceprints. The problem is that if during the model training it didn't see enough faces of a particular race, it will show low accuracy and I don't see how you can improve it without retraining the model. Also, have you tried the InsightFace https://github.com/exadel-inc/CompreFace/tree/master/custom-builds custom build? It should show better accuracy on Asian race (Still not the best)

— Reply to this email directly, view it on GitHub https://github.com/exadel-inc/CompreFace/issues/730#issuecomment-1044664277, or unsubscribe https://github.com/notifications/unsubscribe-auth/AXLJ3D22F5VYKIMCAQJEHXLU3ZNQPANCNFSM5OXSC6NQ . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

You are receiving this because you authored the thread.Message ID: @.***>

[pospielov]

MobileNet is just another model of InsightFace library, it runs faster than usual models, but is not so accurate. Have you tried SubCenter-ArcFace-r100 model? Did it work for you?

[pospielov]

what yml file do you have? The models look like this: https://drive.google.com/file/d/1ltcJChTdP1yQWF9e1ESpTNYAVwxLSNLP/view?usp=sharing

[SarimZH]

Hi pospielov, i did try it but there are a few issues that came along. The docker is consuming more memory than it should and it hangs the system overall. Im using it on Ubuntu 18.04 NUC (hardware: intel core i3 with 4GB RAM)

[pospielov]

Unfortunately, I can't do anything with memory consumption now

[dingyaguang117]

For Asians, 1.1.0-arcface-r100 works fine for me.