2 Sep 2019, 09:20

Computer Vision: A distinct look into our future

Over the past decade, computer vision has had a tremendous impact. From self-automated cars to space exploration, its various uses in surveillance and image analysis has bettered the world. The in-depth analysis of video footage and images using deep machine learning models has become a powerful tool that most companies and even governments are taking notice of.

The main uses of Computer vision include:

Recognition and tracking - the computer tries to link any pattern or feature in the image to a previous feature or activity in its database. Through this

the computer is able to answer questions such as what the object in the video or image is, what it’s doing etc.

Assessment - in assessment the computer obtains features of the object.

such as whether the object is safe, etc.

Prediction - lastly it predicts possible behaviours of the object, and how it will react in the future.

Computer vision has numerous applications in the real world from medical imagery to road safety. The ability to collect high quality data is crucial for the success of computer vision. New innovations have made collecting data more effortless. The drone for example, which has increased the quality of visual data extensively, has enabled data to be visualised from a wide variety of angles enabling more data to be collected.

Who wants to be a Computer Visionist?

If you are interested in learning more about computer vision make sure you attend one of the upcoming Zindi MIIA (Machine Intelligence Institute of Africa) Machine Learning Hackathons taking place on:

  1. 3 September 2019: AI Summit at Leaderex at the Sandton Convention Centre in Johannesburg.
  2. 4 September 2019: AI Expo Africa at Century City Convention Centre in Cape Town.

Contact us at zindi@zindi.africa for more details.

If you are tackling a computer vision problem, here are tips and tricks when solving a computer vision challenge from Obins Chaudhary who will be the mentor at the Leaderex Hackathon:

1. Understand data first. What is provided, what do you think is missing? and what is asked?

2. Understand data processing and feature engineering: even though computer vision techniques do not require manual processing but it gives significantly better results after reducing noise in data.

3. Data annotation and augmentation: it’s always good to look first into data labels and if external data sources can help. How given data can be augmented to extrapolate it?

4. Plan your approach: always start with the simplest methods even if they give you low scores to get a better understanding of the data. Track scores where simple models are performing and failing. So you can focus on specific cases.

5. Know your strengths. What you can build immediately versus build after some learnings. Don’t waste time learning and implementing some very advanced techniques in the first go. Simple implementations will give you hint of which advance technique to focus on.

6. Utilize Zindi forum as much as you can to reach out community and experts.

7. How can you utilize transfer learning? Some of the best models and their weights are open sourced and can be very well utilized.