Deep Convolutional Neural Network for Large-Scale Scene Classification

Background

GatorVision, the team representing the University of Florida, participated in the ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) held as part of the ImageNet and Microsoft Common Objects in Context (MS-COCO) joint workshop organised at the prestigious 2015 International Conference on Computer Vision (ICCV). I was part of the GatorVision team alongside 5 other teammates, led by David Ojika.

Problem Statement

Using the Places2 dataset provided by the organisers of the ILSVRC, the aim was to develop an algorithm for solving the problem of “scene classification“, i.e., the accurate classification of images of everyday objects.

My Contributions

This was the first semester of my Master’s degree, so it was an exciting phase of learning!

Over the course of the efforts, I worked on:

Learning the basic fundamentals of machine learning (ML), model training/testing, among several other topics.
Gaining a deep understanding of the basics of deep learning, convolution neural nets, image processing, among others, through research papers and academic coursework.
Learning the basics of Caffe and GPUs for iteratively developing the neural network models (in collaboration with other teammates).
Understanding the standard ML pipeline: from data preprocessing, feature engineering, model training & testing, to model selection and evaluation.
Performing hyperparameter experimentation on learning rate, kernel size, convolutional layers, and total fully-connected layers, and measuring model accuracy & loss.
Preparing the final poster that was presented in the workshop (led by David Ojika)
Iteratively improving the quality of the poster by frequently incorporating feedback from my supervisor and the team advisor

GatorVision at the MS-COCO Poster Workshop at the International Conference on Computer Vision (ICCV), 2015.
(Click on the poster for more details.)

Impact

GatorVision team secured the 20th rank in the ILSVRC globally.
GatorVision was invited to present the poster at the MS-COCO workshop in Chile and share the approach with several dignitaries of the field.
The GatorVision poster remains permanently publicly available (online) on the prestigious ImageNet

Project Links

Link to the Paper: ILSVRC @ ICCV 2015