First DSTA Brainhack Discussion

The conceptualization of “No Idea!”

“Passionate” was how I would describe the discussion

On a friday evening, a normal person would be chilling at the movies, having a long walk, or even clubbing the night away.

It isn’t so for some geeky nerds. The 4 man team which consists of one of my secondary school friends, and two others from my previous hackathon, HackOMania 2024, engaged in a hearty discussion and an exchange of ideas on the upcoming “Brain Hack” hackathon by the Defense Science and Technology Agency (DSTA). We chose the theme, “Today I Learned AI” .

The problem statement, or mission, DSTA calls it.

For my friend, he is more inclined with hardware, as with one of the teammate whom I was acquainted with from HackOMania 2024. They discussed about the DJI robot and the hardware implementations that were possible, such as mounting a separate module and a laser to accurately detect the specified object.

As I lean more towards the software side of things, I presented a model which leveraged RetinaNet with a ResNet50 backbone and Feature Pyramid Network (FPN) for object detection. The reasons this is chose is that, after much research and trial and error, RetinaNet works best for our current dataset where there are a large number of classes and small objects (Aerial view of aeroplanes). The FPN helps in handling objects consisting of different scales effectively and constructs a multi-scale pyramid of feature maps which allows the model to detect objects at different scales and aspect ratios.

Although this seems good, it is only good for viewing at this particular angle, and it is far from complete. There are some weaknesses I have identified such as,

1. Viewing angles may cause the size of aircrafts or objects in the air to look smaller or bigger than normal, thus, I will have to train the model with various rotations and orientations of a given aircraft.
2. Weather conditions. During rainy days or when the sky is filled with thunderclouds, these play a part in obstructing the identification of objects.
3. Identification of friendly and non-friendly aircrafts. In a war setting, there might be scenarios where there is an aerial engagement of various aircrafts. The turret must identify which aircraft is friendly, and which aircraft it must shoot down.

And some others that I am sure I have not thought of.

All in all, it was a good and defining start to further discussions towards the eventual competition, and who knows, maybe we can bag a spot in the finals? *wink*