Inspiration
The challenge of solving real world problem in such a short period of time, while also enduring the multiple emotions during the project. Where at time everything appeared to go wrong, to finally be able to tackle each problem one by one
What it does
Our project takesin images as input cleans them with a Denoising model, enhancing image quality, feeds them to a clasificator model which then outputs a result based on if there is people in the image or not.
How we built it
We built it using a data pipeline and generating multiple datasets which we then we were able to tailor for our specific models, and use them as benchmarks in later steps.
Challenges we ran into
Our choice of the classificator model prove to play anegative role in the overall performance of the entire ensamble model. Due to the time constraints we were only able to focus in the Denoising model which then of course led to us having a tight schedule with very little time to properly test and train our models. Furthermore in the early ours in the morning at around 1-2am our aws account stopped working putting extra constraints on our computing power forcing us to train locally, which made us loose precious time.
Accomplishments that we're proud of
Our denoising model proved to be very effective at removing occlusions from the images, our improvement in the benchamarks though marginal showcased that this approch is worth considering and developing further.
What we learned
Time management, clear communication and clear prioritization of tasks to use the little time we had as effectivly as possible.
What's next for Denoise2Detect
Impvoing the clasiffciation model, as well as the pipeline to make it more dynamic. With more time we could have also implemented more tailored solutions with our models instead of relying on very rushed constrained training loops. We also had to abandon a feature which would have brought great business value. A metadata aggragator which would be able to add features into the metadata of the picture about the scene, the quality of the picture, etc. Due to time contrainst we were unable to implement it, but we would love to add it in the future.
