This assignment deals with creating an ML model which when provided with an input image will accurately describe the objects present in it along with associated characteristics. You should also identify the main object in the image which would be the item to be sold, and try to describe it with the color and other features. Let us go through a sketch of the process:-
-
Annotated Images: The ML model needs to be trained by providing it with a good quality dataset of images mapped to a description of the objects present in it and marked by boxes to show where they occur within the image. Ensure that the dataset contains a wide variety of images with the objects occurring at random locations.
-
ML Framework: There are several prebuilt ML libraries that can be used to develop object detection models without starting from scratch. These libraries powerful tools and functions for training ML models. These libraries come with pre-implemented models, including ones for object detection like YOLO (You Only Look Once). An ML framework can be setup locally by installing all the necessary libraries associated in Python.
-
Configuring the Model: Based on the framework chosen, one can start from scratch or utilize a pre-trained model and modify it as per the need. In the case of YOLO, one can provide the dataset to a pre-trained model and observe the various changes in accuracy and average precision. This can be used to tweak the weights of the model.
-
Testing: Create a testing dataset with ground truths similar to what the model might encounter once deployed. Poor performance on this dataset might indicate a need to change the model architecture.
-
Deploy: Once, the model performs satisfactorily in the testing phase, we can move to deployment. This will involve migrating the trained model from your local setup to a remote environment allowing users to access it.
-
Improve: The weights of the model can be adjusted periodically by utilizing live data itself. Automating the training and deployment cycle with some intervention will greatly improve the model.
-
Image Data Collection and Processing
-
Machine Learning Frameworks for Images
-
Live Deployment/Demonstration