Person detected (92%) in error


#1

Hello, I think the model needs to be tweaked a bit as I keep getting high confidence matches for people that don’t even resemble a human object (see attached example).


#2

Hello @sansnil, thanks for reporting this. You can run in HIGH mode for better accuracy, we are also adding support for finetuning and deploying the detection models on your own dataset. This would be announced in coming updates.


#3

Hi @john , thanks very much for the response. I made the suggested change to HIGH mode and it did reduce the number of false positives for sure. There are still some which occur. What’s the best way for me (or anyone) to provide feedback with the goal of improving the detection model. I’m pretty new to DeepStack but one of the attractions for me is the local installation which I strongly prefer. But it does occur to me that one downside may be there’s no way for all the users to help ‘train’ the model over time.


#4

Hello @sansnil we understand the need to improve the detection as well as add support for detecting objects more than the current 80 classes that DeepStack supports. That’s why we are working on adding the ability to train and detect your own objects easily.
To this end, we recently released a new version of our open source toolkit for training new object detection models, see

The next updates to DeepStack would allow you deploy these custom object detection models easily.