Custom model for recognizing people


I want to build a custom model to recognize only people in a very “noisy” environment, where using the built-in model results with many mistakes (i.e. identify non-people objects as people).
Should I tag on my dataset only people, or should I tag as many objects I can? Preferably the objects wrongly identified as people?

Any tip or recommendation will be appreciated.


For false positives in person detection, DeepStack will most likely report a lower confidence score than true positives.

You can filter detections my setting min_confidence to values between 0-100 in your request to DeepStack ( e.g min_confidence=70 ) will ensure false positive detections are filtered from the response.

Thanks @OlafenwaMoses, but in my case the false positives range from very low to very high confidence values. Anyway, I’ll think on a different approach to solve the problem…