Challenge Unboxed📦: AI Blitz⚡8
Introducing Challenge Unboxed: AI Blitz 8 📦!!
Through this series, we delve deeper into the unique heuristics and methods used by winning participants in our challenges. This series aims to uncover new tools and approaches that will help you broaden your horizon in Machine Learning.
In this segment, we look over the methods leveraged by winning participants in the AI Blitz⚡8. We shall probe the solutions that scored the highest on the leaderboard. Don’t forget to add these tools and hacks to your ML arsenal.
👀About the Challenge
AI Blitz⚡8 is our monthly 21 days long AI challenge that includes 5 exciting beginner-friendly AI puzzles. Blitz 8 was themed around the world’s popular sport of Formula 1.
The challenge presented participants with computer vision problems in the context of an F1 race. With participants roleplaying as managers for their F1 team and coming up with innovative solutions, the end goal is to get the highest score in all 5 puzzles to win an exciting cash prize.
Let us look at the puzzles presented by the challenge:
Dennis Tsaregorodtsev topped the AI Blitz 8 leaderboard! We also saw previous Blitz winner G Mothy share exceptional implementation for a puzzle. Their submissions stood apart from the others as their solutions demonstrated innovative thinking and rigorous research. Let us look at how they solved the problems.
🚭Smoke Elimination using GCANet
The “F1 Smoke Elimination '' puzzled was labeled as “the freshest CV problem” with a lot of Blitz users enjoying its complex nature. Participants were required to eliminate smoke from frames of cars in accidents. Due to the uniqueness of the problem, participants were compelled to look into the advanced Deep Learning algorithms for Computer Vision.
One such approach was the use of Gated Context Aggregation Network for Image Dehazing and Deraining or GCANet by the challenge winner Dennis. In his submission, he treated the prompt as a dehazing task. Instead of implementing GANs like many other participants, Denis found the GCANet approach by conducting a literature survey.
GCANet adopts an approach where the architecture uses smoothed dilation technique to the convolutions to not only better the previous at the time State-of-the-Art but also removes considerable redundant variables. Contrary to how in traditional approaches, it has been noted that CNNs tend to work better with Dilation introduced to them, i.e. convolution applied to input with defined gaps. However, Dilated Convolutional Networks tend to introduce Gridding Artifacts or anomalies in the output which eventually hinder the performance of a Dehazing Network.
However, the SOTA architecture for the task has now changed to FFANet which one may try for a better solution. What method would you try for dehazing an image?
🏎Speed Recognition using Ensemble Learning
Racing at unreal speeds, F1 car drivers always need to be aware of their numbers. The puzzle “F1 Speed Recognition” asks participants to recognize the speed of a car by glancing at the fault speedometer.
The puzzle calls for a well-optimized and efficient digit recognizer that’s able to recognize numbers regardless of the background color. This problem attracted a rather interesting approach from G Mothy, the runners-up for the challenge.
In his submission, Mothy implements the following steps:
- Use clustering to get the type of speedometer.
- Make the pixels around the speed into white color to localize the text.
- Apply OCR to detect the text from the images.
- Use Image Regression to train the model.
- Use Ensemble Learning on the OCR output, Image Regression output.
The key step that stood out in his approach was the implementation of Ensemble Learning for the actual Recognition task of the extracted digits. Ensemble learning is the process by which multiple models, such as classifiers, are strategically generated and combined to solve a particular computational intelligence problem.
For his implementation, Mothy decides to go for ResNet18, VGG19, and ResNet50 as the models to the ensemble as his feature extractor. What CNN architectures do you think would have worked better, let us know in the comments?
👓F1 Car Detection using Faster RCNN
Object Detection has been observed to be at the center of Computer Vision prompts with the sudden influx of automation projects. One such application that was introduced to the participants in the form of a puzzle was, “F1 Car Detection”. In case of accidents, the first and foremost task that remains by the rescue is to localize the area of accident to which Object Detection acts as the savior.
Contrary to traditional Object Detection methods using Segmentation by Region-Based Convolutional Neural Networks or RCNN, G Mothy, the runners-up for the challenge, was seen using Faster RCNN in his implementation.
Working on a VGG CNN architecture backbone, Faster RCNN introduced a Regional Proposal Network which is a fully convolutional network that simultaneously predicts object bounds and objectness scores at each position. When working at higher efficiency Faster RCNN is able to give Object Detection at about 10 FPS compared to 5 FPS from Mask RCNN, Faster RCNN’s predecessor.
Which challenge would you like us to Unbox next? Comment down below or tweet us @AICrowdHQ!