I did it
At the time I’m writing this article, I have already passed this semester including Thesis with a perfect score on my presentation. In the following weeks I’ll prepare a more detailed article on my topic and everting that I presented.
For now I will show you some of the changes I did since last time and I show the results.
If you’d like to now more about the context of my research you could watch the other articles; however, here’s a small recap. So I wanted to work with a fall detection system, how ever there are many moving parts and my advisor was able to guide me to a more suitable approach for an undergrad thesis project which was presenting a new framework based on Siamese Neural Networks that used Transformer Encoders rather than CNNs to detect fall events.
Not only did I compare the performance of CNNs but I compared the implementations of three different transformers.
The Poolformer from the research “Metaformer is actually what you need for vision”
The “Classic ViT” from “AN IMAGE IS WORTH 16X16 WORDS”
and an extension from the ViT using Shifted Path Tokenization and Locality Self Attention
These three transformers in a Siamese configuration were compared with MACs which is a measure for computational complexity and Accuracy. The images below show the results.
URFD dataset
FDD dataset
Clearly any transformer implementation beats the CNNs implementation by far. I’ll talk more in detail in a future article. Thank you very much for your attention :)
