Transfer Learning involves storing knowledge gained while solving one problem and applying it to a different related problem. In our case, knowledge gained in learning person, cars etc can be applied to detect PPEs.

The advantages of transfer learning are:

1. Less Training Data

2. Improved Neural Network Performance

3. Reduced Training Time

Faster-RCNN was originally proposed in 2015, as the third iteration in the RCNN family of models for object detection.

It's architecture involves:

• RPN - generates Region Proposals i.e where to look in the image.

• RoI pooling layer generates a feature vector for each region.

• Fast-RCNN layer classifies the feature vectors and return class scores and bounding boxes.

For our project, we did the following:

• Experimented with 2 Faster-RCNN model with ResNet50 and MobileNetV3 backend .

• Training with SGD optimiser with momentum 0.9 and learning rate 0.005 for 100 epochs

We obtained the loss curves as below. The steep decrease in loss suggests that transfer learning works well for our dataset.

YOLO (You Only look once) was Published in 2015 . Convolutional Neural Network based Object Detector.

• It had several revisions for increasing speed and performance

  • First, Yolo-v1 to latest Yolo-v5

• YOLO divides an image into several grids and each grid cell predicts bounding boxes and confidence scores for these boxes.

YOLO's architecture comprises of:

• Model BackBone ( CSP in Yolo-v5)

  1. Extracst rich informative feature from input image.

• Model Neck ( PANet in Yolo-v5 )

  1. Generates feature pyramids for scale generalisation.

• Model Head ( Yolo )

  1. Final Detection part

  2. Output boundary boxes and class probabilities