A quick guide on installation of important libraries and running the code. The project has four .ipynb files - watch_classification.ipynb, regression_watch.ipynb, bags_classification and bags_regression.ipynb
Both of the classification and Regression ipynb files consists of 3 models each namely a Custom Model, InceptionV3 Transfer Learning model, DenseNet201 Transfer learning model. Thus in a way we trained 12 models for this project.
We have built a streamlit application which takes an image input of a bag or a watch. The website then predicts the brand and the estimated price of the product given as input. Note: Our models take a lot of space and hence the web application is about 2 to 3 GB. We are facing some storage issue while deploying it currently
To create efficient models , it is important to have good quality data. Maybe more in-depth pre-processing techniques could help improve the model performance.
- Image Data Generator helps reduce overfitting our data.
- Using complex models will improve performance but needs more computation power and might not always be the best solution.
- Image regression for price prediction is a difficult task when you have many unique price points. Regression may be improved if we put prices range in various categorical bins.
© Sameer Mahajan © Harsh Deokuliar