This repository is to build a model that detects people, heads, and hardhats in images. Each image can contain multiple people, multiple heads and multiple hardhats. The solution will help build a safety app that will alert the user if there are workers in the field that do not comply with safety rules. This Notebook is for Training Purpose of Keras RetinaNet. RetinaNet is very slow as compared to F-RCNN so I've kept epochs and steps per epoch small for fast commiting purpose.
Step-1 : Installing Keras-RetinaNet Step-2 : Let's look at the data Step-3 : EDA Step-4 : Visualizing images
What can we tell from visualizations:
there are plenty of overlappind bounding boxes all photos seem to be taken vertically all plants are can be rotated differently, there is no single orientation. this means that different flip and roration augmentations should probably help colors of wheet heads are quite different and seem to depend a little bit on the source wheet heads themselves are seen from very different angles of view relevant to the observer
Step-5 : Preprocessing Data for Input to RetinaNet
Step-6 : Preparing Files to be given for training Annotation file contains all the path of all images and their corresponding bounding boxes Class file contains the number of classes but in our case it is just 1 (Wheat)
Step-7 : Downloading the pretrained model
Model Parameters