Wondering what it’s like to work on an Omdena challenge? Are you unsure of how you will be able to contribute? Don’t worry! Just walk through the looking glass (this article 😄) to enter the amazing world of collaboration and learning.
By Sneha Bahl
The Omdena-Weedbot project was my first collaborative project. I was excited to work with new people and grasp as much as I could working on this real-world problem. I was also nervous, not completely confident if I could contribute enough (Don’t worry, this is completely normal). The goal of the project was to be able to detect weeds amongst the carrot crops. The objectives and data were provided by the Weedbot team.
The goal is to develop a high-speed plant image recognition neural network with a speed of 12ms per image or faster and recognition precision of 100–110% of crop polygon, which means up to 10% false positives are allowed.
“What it’s like to work on an Omdena challenge”
What’s a better way to understand what your journey will be like in an Omdena project than to actually walk through the whole 8–10 weeks you’re going to be involved in?
As I mentioned before, this was my first Omdena project. I was thrilled as I was going to be a part of a team of 50 collaborators from all around the world working towards the same goal.
Within the first week, I started to feel overwhelmed by seeing everyone participate and share many resources and I had no idea where to start from. I tried to go through the material shared by Omdena and the Weedbot team to get the objectives clear. Attending the Zoom meetings gave me some hints where to start, and overcome the clueless state I had.
During the second week, the team started to create tasks and start working actively. I was starting to feel a little anxious as I was not able to make any meaningful contribution and even catching up on resources was getting harder. I wanted to be a part of all the tasks to maximize my learning.
In week three, I started working on the Feature Generation task and started exploring some new classification and segmentation models we could try and experiment with.
By week four, I realized I was trying to be a Jack of all trades and master of none. I started focussing on a lesser number of tasks and thus being able to contribute more efficiently. Through Slack channels and other collaborators’ help, I started to think clearer and concentrate on delivering solutions.
In week five, I was confident enough to get the responsibility of being the task manager for the Data Annotation task midway. The goal of this task was to change the annotations of plants from single class to multi-class. We believed through reading research papers and many articles on the web that this approach would help the computer vision model learn better.
By week six, we changed the annotations of over 10,000 plants. Since multiple people were working on this task, we decided to review each other’s work and found a lot of inconsistencies. We brainstormed through sprint calls on how to tackle the problem and came up with a solution to create two new classes which could handle the border cases and another round of review. After this we found the annotations to be much better and consistent.
Around week seven, we performed the initial testing using this multi-class data we realized this approach did bring improvement in the model performance. We added two more goals for this task: annotating new images from scratch which were provided by the Weedbot team and cropping images to a smaller size. At the first look, cropping looked to be an easy job but there were a lot of things to take care of.
Challenges in cropping images
- We started with the most basic approach of just cropping the images of initial size 3008 X 3008 to a size of 1920 X 550. On visualizing the cropped images we realized that a majority of crops were cropped midways, hence changing their appearance. A crop that was initially annotated as class1 now appeared to be of class2. We thought of fixing this manually but this was a very strenuous task as it meant to change annotations in over 10,000 images.
- We tried a different approach where we took the maximum size possible of a crop and chose the cropping size to be 1920 X 1200. This did reduce the frequency of cut-out plants but still, there was a considerable number of such cases.
- Lastly, we tried an adaptive approach, where we considered the position of crops in the image and decided the cropping window based on that. We also used a background mask to cover any plants which were less than 80% within the cropping window. This approach worked for us perfectly.
Using these two ideas- multi-class and cropping of images brought around 5x improvement in the mAP (Mean Average Precision), as compared to single class non-cropped data.
Wrapping up all the hard work we have accomplished started so I volunteered to work on setting up the documentation and work on the Final Presentation. This helped me go through the whole project deeply, learn about the tasks I was not directly involved in, and reflect on what all things worked and what didn’t.
I remember being skittish while briefing during the first weekly meeting, but by the end of the project, I was much more confident and was able to deliver well in the final presentation. That’s what they say about Omdena’s challenges, you finish them with a new handful of senior skills.
Some lessons I learned
- Get out of your comfort zone: This is very important if you want to grow and learn. For the first few weeks, I was not confident of being able to contribute anything to the project. But once I pushed myself and volunteered to work on a task, others helped me and guided me on how to go by it.
“Life Begins At The End Of Your Comfort Zone. ”
— Neale Donald Walsch
- Be patient: Working with people from different timezones helped me learn that one needs to be patient and everyone needs their time and space to work. Working with a team of 40 people is much harder than working with a smaller group. One needs to be considerate and understanding of others’ availabilities and commitments.
- Every contribution counts — No contribution is big or small, every contribution counts. While working on annotating new images, we realized it was a time taking task and we needed manpower. Many collaborators reached out o help with the task and by just giving their 2–3 hours they saved another person around 10–12 hours of work.
Why should you work on an Omdena project?
The benefits of working in a collaborative environment like this:
- You get to learn a lot! Everyone has something to offer be it technical skills or management skills or soft skills.
- Networking: This is a great way to make meaningful contacts and build your network.
- Community: You can be a part of the Omdena community and continue to contribute to solve real-world problems and use AI for Good.