Computer Vision Workbench

  • Machine Learning

  • Image Processing

  • Development

  • Deep Learning

Mentors :

  • Vishal Kaushal

Mentees :

  • Aniket Sadashiva

Extracting potentially useful information from videos, like presence of faces, humans, specific objects, motion, tracking etc. is an indispensable part of video analysis.

Recent advancements in deep learning have demonstrated good accuracies on many of these tasks, however these models do not yet generalize well to the unseen real-world deployments.

Though the accuracy numbers for these off the shelf models have been reported and can be computed on standard existing datasets, there is currently no way to estimate how would these models perform when challenged in real-world deployments. This makes it necessary to have at least a qualitative comparison of different models against each other on such challenging, unlabelled videos.

To enable this, we envision a workbench tool (Python GUI) which will make it easy to compare any two models on a given video for a given task and “see” how they perform. It will be released as open source software. This tool will help understand the “real” strengths and weaknesses of different models for different tasks and will help give important directions to undertake future research.

To begin with we can focus on the following computer vision tasks: Motion detection, Face detection, Face recognition, Human detection, Head detection, Object detection, Tracking. A typical use of this tool will involve selecting the task, selecting a video, selecting two models/algorithms to compare and as a result seeing a side by side comparison of the analyzed video played synchronously along with other measurable parameters like time taken etc.

Tentative Timeline :

Week Work
Week 1 Understanding the various computer vision tasks involved
Week 2 Motion Detection - stand alone implementations using various alternate techniques
Week 3 Face Detection - stand alone implementations using various alternate off the shelf models/techniques
Week 4 Object detection - stand alone implementations using various alternate off the shelf models/techniques
Week 5 Tool design, architecture and skeleton implementation
Week 6 Integrating motion detection in tool
Week 7 Integrating face detection in tool
Week 8 Integrating object detection in tool
Week 9 Face recognition
Week 10 Human detection
Week 11 Head detection
Week 12 Tracking
Week 13 Buffer/Enhancements/Closure