quicktake 0.0.16

Off-the-shelf computer vision ML models. Yolov5, gender and age determination.

QuickTake

Off-the-shelf computer vision ML models. Yolov5, gender and age determination.

The goal of this repository is to provide, easy to use, abstracted, APIs to powerful computer vision models.

Models

$3$ models are currently available:

• Object detection
• Gender determination
• Age determination

Model Engine

The models

• YoloV5: Object detection. This forms the basis of the other models. Pretrained on COCO. Documentation here

• Gender: ResNet18 is used as the models backend. Transfer learning is applied to model gender. The additional gender training was done on the gender classification dataset, using code extract from here.

• Age: The age model is an implementation of the SSR-Net paper: SSR-Net: A Compact Soft Stagewise Regression Network for Age Estimation. The pyTorch model was largely derived from oukohou.

Getting Started

Install the package with pip:

pip install quicktake

Usage

Build an instance of the class:

from quicktake import QuickTake

Image Input

Each model is designed to handle $3$ types of input:

• raw pixels (torch.Tensor): raw pixels of a single image. Used when streaming video input.
• image path (str): path to an image. Used when processing a single image.
• image directory (str): path to a directory of images. Used when processing a directory of images.

Expected Use

Gender and age determination models are trained on faces. They work fine on a larger image, however, will fail to make multiple predictions in the case of multiple faces in a single image.

The API is currently designed to chain models:

1. yolo is used to identify objects.
2. IF a person is detected, the gender and age models are used to make predictions.

This is neatly bundled in the QuickTake.yolo_loop() method.

Getting Started

Launch a webcam stream:

QL = QuickTake()
QL.launchStream()

Note: Each model returns the results results_ as well as the runtime time_.

Run on a single frame:

from IPython.display import display
from PIL import Image
import cv2

# example images
img = './data/random/dave.png'

# to avoid distractions
import warnings
warnings.filterwarnings('ignore')

# init module
from quicktake import QuickTake
qt = QuickTake()

# extract frame from raw image path
frame = qt.read_image(img)

We can now fit qt.age(<frame>) or qt.gender(<frame>) on the frame. Alternatively we can cycle through the objects detected by yolo and if a person is detected, fit qt.age() and qt.gender():

# generate points
for _label, x0,y0,x1,y1, colour, thickness, results, res_df, age_, gender_ in qt.yolo_loop(frame):
    _label = QuickTake.generate_yolo_label(_label)
    QuickTake.add_block_to_image(frame, _label, x0,y0,x1,y1, colour=colour, thickness=thickness)

The result is an image with the bounding boxes and labels, confidence (in yolo prediction), age, and gender if a person is detected.

.

The staged output is also useful:

.

For a more comprehensive example directory.

Future

I have many more models; deployment methods & applications in the pipeline.

If you wish to contribute, please email me @[e-mail].