Updated: Feb 11
Getting the internship:
It was during the month of July when my uncle received a message on a WhatsApp group about a startup in Pune which wanted to hire an intern for machine learning . My uncle asked me if I would be interested in this and I immediately said “Yes!”. Soon after this, my father and I got on a call with Ashish sir, the founder of the Puneri startup - TechArtisan 3D Solutions. We discussed the details of the internship and Ashish sir said that he will review some other candidates and get back to me. The next day, I got a call from him and he said that he wanted to choose me as the intern! I was extremely delighted on hearing this news. He said that this is the first time we are hiring such a young intern but seeing the projects that I had worked on, he said that he was confident I would be able to do it!
Details of the Internship:
The internship was structured by Ashish sir in order for me to balance both my studies and the internship. We would communicate over weekly phone calls and emails and discuss the progress of the projects. Apart from this, the internship would last for 4 months through till November. Furthermore, he would also be paying me a monthly stipend:)
As part of the first project, I had to detect stains on clothes. For humans this task would be extremely simple but for computers this task isn’t as easy as it looks. To do this I used a publicly available dataset which contained images of cloth stains as well as locations in the image where the stain was present. Using this data I trained a custom machine learning object detection model based on the YOLO-v4 architecture and it achieved a very good accuracy. The results from the test set are shown below.
Code to this project:
The first project was a simple one as I had implemented object detection algorithms earlier as well. The 2nd project was an image segmentation task. The goal was to use an existing model architecture and segment parts of an image into different categories. This project would help me do future custom segmentation tasks. To do this segmentation, I used the MaskRCNN model and the results are shown below.
Code to the project:
After having learnt how to implement an already existing segmentation algorithm, my next project involved a custom segmentation model which could segment cell nuclei from an image. Having no previous knowledge about image segmentation, this project helped me learn about several image segmentation model architectures such as the U-net architecture etc. I then implemented this U-net architecture to segment cell nuclei from the images. The model trained very well and received an accuracy of ~98% on the test set. Below are the results of the segmentation task.
The top image is the test image. The bottom image is the segmentation result.
Code to the project:
For this project, we used data from CT scans of patients who had been infected with covid-19. Using this 3D data from the scans, we had to segment different parts of the scan where an infection was detected and found on the lungs. Below are the results of this project:
The leftmost image is a section from the CT-scan. The middle image is the actual segmentation. The rightmost image is the custom model segmentation.
Using this data, we also created a 3D point map for the lung boundaries in each scan. Below is the result of this boundary detection:
The overall experience in this internship was absolutely splendid! Not only was I able to apply my knowledge in ML but I was also able to learn new model architectures and solve new problems! Thank you TechArtisan and Ashish sir for giving me this opportunity. I will always cherish this experience of my first internship!