# Compare Asian Celebs model with official R50 MxNet InsightFace

## Method
Use Jetson Nano to recognize the faces from webcam with resolution 720x1280 then analyze and compare the result.

Data is collected on 6 days.

The webcam is installed in front of HR room (Neolab Vietnam) and it always sees `Truong Phan Hoang My`, so the number of faces of this employee is very big (compare with the other employees) so I will ignore this employee so the data doesn't screw.

Most of the unknown faces also belong to My-san with difficult pose of face or bad light condition.

## Result
https://drive.google.com/file/d/1J5eNGEVzXa1mkaFp1epVyt9AcMj5p4hr/view

## Some comments on that result
* Both models has its fail cases, especially with bad image quality (blur and dark)
* The official model can recognize more people than the model that trained with Asian Celebs dataset
* The number of **unknown** faces is also smaller in official model.

## Conclusion
* At current step, the official model is better than the model that trained with Asian Celebs dataset
* The model that trained with Asian Celebs dataset is just a checkpoint durring training process, it's not converged yet, so it's not the final version. We should continue training it.

## Some other note about the quality of camera
I compared the faces that captured at Neolab VN office to the faces that capture JP side and found that the image quality of JP side is much better. I think it's because of the webcam and also the light condition.

I calculated the brightness level and the blurness level of both.
* Darkness value runs from 0 to 200, 0 means very dark and 200 means very bright. The good value should be in the middle (around 100, I thought)
* Blurness value: we usually consider a picture is blur if blurness value < 100, and not blur (sharp) if value >= 100
* Maybe it's because the webcam can't auto focus, so the faces are not the main target of the frame and it's blur.

For JP faces (that Naoto-san sent me):
* Average blurness: 167.9 (not blur)
* Average brightness: 112.2 (good)

For VN faces:
* Average blurness: 66.6 (blur)
* Average brightness: 64.9 (dark)

The current webcam on VN office: I checked (by Linux command) and found that the model is Logitech B525: https://www.logicool.co.jp/ja-jp/product/b525-webcam?crid=34 but the real device looks different (on logo)

Here is the detail different about image's quality between JP and VN side




## Next step
* Continue to train with Asian Celebs.
* Calculate blurness and brightness, ignore the bad faces before recognizing.
* Try to pick good camera.