about nasa comet detection marathon

 This was an interesting comet detection task on topcoder. The main challenge was to detect very faint comets on LASCO C2 telescope images.

I took prize 7th place.

My approach was to train retinanet on resolution 1024x1024 using as an input - concatenation of 3 sequential differences of images using combined bounding box for 3 sequential comet positions.

The augmentations were used: rotate90,horizontal/vertical flip. 

The provisional score: 38.04.

After detection inference - I used postprocessing - tried to find max group of points which lie close and form a line - and then I tried to extend this group by time linear interpolation assuming constant speed.

What not much helped:

- for preprocessing - I tried histogram equalization - it helped visually better to see faint objects but for solution it did not helped much.

- for preprocessing - I also tried bilatel filtering - but it also didn't much help.

- deepsort (the trajectories are mostly linear, one comet per sequence)

- false positive filtering  (filtering FP by month of year - its known that at month m - there can be trajectories of specific form and orientation)

What could help and I've not tried:

- modeling comet trajectories as curved (not straight lines)

- ensembling

- heuristic way to constructing trajectories (as was done in asteroid hunter contest 2014)