How to overcome double-counting in object detection / Non-Maximum Suppression

[Evan Vaughan]

We are currently using the Object Detection feature with and have run into an issue with double-counting. I’ve added a hand-drawn illustration to show the issues we’ve been running into.

We have tried to mitigate the issue by

1) using a confidence threshold (we’ve found .65-.70 works well for our purposes)

2) using a time threshold (i.e. object must be detected for longer than 0.25 seconds)

3) using if/else statements taking into account the IoU (this can get sloppy and we haven’t been able to get it to work with success across the majority of our videos. What works for one video doesn’t necessarily lend itself to other videos.

Has anyone come across the same issue - and I was wondering if anyone else is having the same issue and if so how they have overcome it.

I was going to try and implement a Non-Maximum Suppression algorithm next but thought that Google might have built it into the model already. Does anyone know if that's the case?

Any help would be greatly appreciated!

Disclaimer: The information in this email and any attachments may contain proprietary and confidential information that is intended for the addressee(s) only. If you are not the intended recipient, you are hereby notified that any disclosure, copying, distribution, retention or use of the contents of this information is prohibited. If you have received this e-mail in error, please immediately contact the sender and delete the e-mail.

[Evan Vaughan]

This is using the VideoIntelligence API

[Jun Lu]

Hi, 

Google Groups are reserved for general product discussion, StackOverflow for technical questions whereas Issue Tracker for product bugs (unexpected behaviors) and feature requests.

To get a better support you should post to the relevant forum, thus please read the Community Support article for better understanding.

On Tuesday, January 21, 2020 at 4:52:30 PM UTC-5, Evan Vaughan wrote: