I created a custom dataset of pistols and rifles. Ran ssd with focal loss but didnt get very good results. Thanks to @sgugger’s implementation of Retinanet which worked pretty well on our custom dataset of only 1043 images out of which 209 were used for validation. I ran this model on a video, sharing the results here . Please give your feedback.
I was able to finish this within 10 months of taking the fastai course thanks to Jeremy. Will post a blog post describing how I went about this project soon.
Personal update: Because of this project I got an invitation to be a Tensorflow mentor for Google Code-in 2019. Thanks to Jeremy Howard Rachel Thomas Sylvian Gugger and the entire fast.ai community