Refactor face extraction and alignment logic; enhance bounding box transformation and facial landmark adjustments

[jogendra-india]

Tickets

#1413

What has been done

In this PR, I have modified the code so that bounding box around image also gets rotated on image rotation during alignment.

Here are some result:
Original Image I used (left rotated):

Here is the earlier aligned image extracted (plotted aligned_sub_img from extract_face() function that is return of align_img_wrt_eyes function)

After the change in PR, this is how it looks now:

Tested the performance on random images to identify the same person using embeddings from the image below.

Image 1:

Image 2:

Image 3:

Here are the results:

Cosine Similarity Comparison Table

Metric	Image 1 (Original)	Image 1 (After Change)	Image 2 (Original)	Image 2 (After Change)	Image 3 (Original)	Image 3 (After Change)
Facenet	0.63	0.65	0.78	0.75	0.48	0.41
ArcFace	0.45	0.39	0.47	0.42	0.29	0.19
VGG-Face	0.40	0.39	0.46	0.42	0.64	0.73
Facenet512	0.69	0.63	0.66	0.58	0.85	0.90

Observations:

1. Image 1 shows a slight improvement in Facenet (0.63 → 0.65) but a slight drop in other metrics.
2. Image 2 maintains overall performance but slightly decreases across all metrics after the change.
3. Image 3 (left-rotated) demonstrates better alignment with VGG-Face (0.64 → 0.73) and Facenet512 (0.85 → 0.90), indicating improved handling of rotations.

@serengil Not entirely sure if this has resulted in significant improvements. While the alignment issue appears to be resolved, there may still be other areas requiring adjustments. I had anticipated cosine similarity improvements across all models, but the results seem mixed.

I would appreciate it if you could thoroughly test this with your expertise and recommend any additional changes that might be necessary.

Thank you!

How to test

make lint && make test

[serengil]

Hmmm, the scale of these changes seems risky given the minimal or negative impact. It is not worth the trouble.

[jogendra-india]

Hmmm, the scale of these changes seems risky given the minimal or negative impact. It is not worth the trouble.

Agreed! Alternatively, we could limit this change to cases where the eyes are positioned vertically, such as in rotated images. However, this might be over-engineering. I'll revisit this later in the week if I come up with other adjustments. How does it look on your end with your test images?

[serengil]

seems it passed the unit tests