I’m totally new to object detection. I want to learning the subject by starting a project [I’ve found in the past such approach keeps me invested & interested].
I’m most interested in achieving very high mAP. Detected objects never overlap, but very frequently are obfuscated, although they always retain their shape. Which avenue would you suggest to pursue? Time is less of an issue, anything less than 1000ms should be fine.
Don’t focus purely on mAP; you should be optimizing for how well your model generalizes to examples outside of your training set (where visual validation can be more useful than training stats).
High-quality data is going to be critical for you, I suggest investing most of your time here (vs model parameters and tuning).
To start, I would go for a one-stage object detection model. In the case you aren’t getting the accuracy you need, you can always train on an aggregated class and pass the model output into a second stage classifier.
Here’s a guide on how to iteratively build models, which you might find helpful.
