Rachana’s Status Report 10/22

This week we had our design report due. While writing out the design report, we were able to resolve many confusions. I mainly worked on training and testing the model with a couple photos. I had to shift from Tensorflow because documentation for Tf was limited, and there were many version conflicts. Upon the recommendation from the TA, I switched the code base to Pytorch tools mainly. I had to spend time removing problematic images as well as they were causing a lot of read errors with the images in the train directory. I am currently able to train and test two classes successfully with a ResNet-50. There are many unused parameters still which is probably eating up a lot of memory that I need to work on fixing. The following week I want to see if I can map out an accuracy vs model test and see which is the optimum model where I optimize the parameters I am using. I also want to figure out a segmentation model where I can create a bounded box around the main object being looked at. I also want to explore how we can best train the model to look at images from both the camera angles -> would it be better to train it with a tuple of images or would it make sense to go for a probability distribution on which camera usually gets the right frame that makes it possible to identify and classify the object. I think I am at a better spot than a week ago. I want to look at having a prototype for UI set up for the website as well by the end of the week.