Optimize ssd detector

[kremnik]

What has been done

With this PR, pandas has been removed from the SSD detector script. All computations are now performed in matrix form using numpy, which speeds up calculations and reduces memory consumption.
Additionally, a test for SSD was added to test_extract_faces.py, and a bug related to checking for a black image was fixed: cv2.dnn.blobFromImage expects np.uint8 as input, while np.zeros by default generates np.float64.

How to test

make lint && make test

[serengil]

IMO, implementation with pandas is much cleaner and readable.

no idea what is going on these lines

        faces = detections[0][0]
        faces = faces[(faces[:, 1] == 1) & (faces[:, 2] >= 0.90)]
        faces[:, 3:7] = np.int32(faces[:, 3:7] * 300)
        faces[:, 3:7] = np.int32(faces[:, 3:7] * [aspect_ratio_x, aspect_ratio_y, aspect_ratio_x, aspect_ratio_y])
        faces[:, 5:7] -= faces[:, 3:5]

PS: would you please raise a ticket before creating a PR to discuss first, otherwise a PR may not be merged, and your effort goes wasted.

[kremnik]

@serengil Ok, I will try to make it more readable.
The purpose of this PR is to avoid using slow pandas when possible. Also, this is not a piece of code that users actively edit.

[serengil]

TBH, I am satisfied with pandas because it comes with readability. No matter users aren't touching this.

[kremnik]

@serengil pandas is great for data analysis and some data pre-processing, but it can be quite slow for production purposes, especially when filtering.
For example, the following code:

detections_df = detections_df[detections_df["is_face"] == 1]  # 0: background, 1: face
detections_df = detections_df[detections_df["confidence"] >= 0.90]

takes around 26ms on selfie-many-people.jpg on my cpu.

However, this code:

faces = faces[(faces[:, 1] == 1) & (faces[:, 2] >= 0.90)]

takes around 16μs on the same photo and the same cpu.

[serengil]

Then, you convinced me to switch :)

But please make it readable with comments at least. Because I have to fight with issues and the understanding the code not written by myself is hard.

Storing index keys to variables would be good.

e.g. idx_face = 1 and faces[:, idx_face] == 1

[serengil]

LGTM