Ravi Theja, Solving Automatic Number Plate Detection and Recognition in India
Ravi implemented NIPS 2017 paper “Selective Classification for Deep Neural Networks”, and is a winner of the Global NIPS Paper Implementation Challenge. See his code implementation here.
Tell us a little about yourself?
I am a Master’s student at IIIT Bangalore with specialisation in Computer Vision. Prior to this, I worked as a data scientist at a hyper-growth startup where I worked on text classification and sales forecasting problems. My research interest lies in the intersection of Computer Vision and Natural Language Processing, including Visual Question Answering (VQA), and the automatic generation of medical reports.
How did you get started in AI?
My journey in AI started in 2017 with taking the Deep Learning Nanodegree offered by Udacity. Later on, I also took up Udacity’s Self-Driving Car Nanodegree (Term 1) to acquire more knowledge and projects to work in the area of Computer Vision.
Initially, I faced a lot of difficulties in understanding the concepts of deep learning especially Recurrent Neural Network (RNN) and Long Short-term Memory (LSTM). I relied on reading blogs that would tear down complex concepts into granular explanations — some of my favourites include Colah’s and Andrej Karpathy’s.
The Global NIPS Paper Implementation Challenge was my first implementation, and it has inspired me to explore new avenues to apply deep learning concepts!
What are you most passionate about in AI?
I am very passionate about solving real life problems using deep learning. For example, I am working with Electronic City Industrial Township Authority (ELCITA) in Bangalore on automatic number-plate detection and recognition.
Even though automatic number plate detection and recognition has already been widely tested and adopted by a lot of countries for surveillance purposes, but in India where the size of number plates on Indian vehicles are not fixed and the CCTV cameras which are used for surveillance purpose are not of high resolution — ANPR remains a challenge to be solved.
Can you give us an overview of your implementation in the Challenge?
In “Selective Classification For Deep Neural Networks”, the authors proposed a new method to build a classifier on top of a model where the classifier selects or rejects a particular instance. If the classifier predicts that the model can predict the instance with great confidence then it would accept the instance and passes it to the model; otherwise it would reject the instance and tells that a human intervention is needed. Consider the example of self-driving cars, the classifier would call for human intervention if the car can’t predict if there was a person on the way with high confidence. This can be used in medical applications in which high precision is critical.
Were there any challenges while implementing your selected paper?
Understanding the mathematics in the paper was the biggest challenge. Also, obtaining benchmark accuracies on Imagenet validation set using pre-trained models. There are a few things I learned from implementing a paper for the first time, which might be useful for some of you who would like to get started in implementing research papers:
- Select a paper according your interest — I am interested in computer vision, so I shortlisted a few papers in the field. While choosing the paper, I looked at how interesting would the future work of the paper be and how I can use the implementation for my future projects.
- Read the paper twice or thrice. I find reading the reference papers cited in the paper helped with my understanding too!
- Find mentors — the expert mentor sessions are really helpful in getting my doubts cleared, thanks to Ritwik Gupta from Carnegie Mellon University and Prof. Srinivasaraghavan from IIITB. I took the chance to reach out to the author of the paper (thanks to Yonatan Geifman!) as well whenever I was stuck.
What’s next for you in your work?
I am currently working on Automatic number plate recognition on Indian vehicles, Necrosis region localisation in histopathological images, and sarcasm detection. Once they are done I am looking forward to work on VQA datasets!
Ravi is a Master’s student at IIIT Bangalore. To keep up to date with Ravi, check out his blog and Github.
This is a feature of the winner of the Global NIPS Paper Implementation Challenge. You can read other winners’ feature here. Let us know if you enjoyed this series and would like to see more of content like this, drop us a comment or an email at info@nurture.ai
