How to keep raw detections for certain classes after tracking

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

In my project there are multiple classes that I need to track but at the same time there are classes which shouldn’t be tracked because otherwise the tracker would discard some detections that are needed per frame.
Is there a way I could flag to not track certain classes and just pass the raw data instead?

4

nvtracker don’t support bypass objects based on class. What is the mean of above? Do you mean the number of the object lager than maxTargetsPerStream: Gst-nvtracker — DeepStream documentation?

5

no, I meant that if the tracker decides that an object is not part of a track, the object will be discarded and only appears if being tracked.

6

There is shadow tracking in nvtracker: Gst-nvtracker — DeepStream documentation if the tracking confidence is low. If so, you can tune nvtracker config parameter(refer: Troubleshooting — DeepStream documentation) to avoid it instead of bypass nvtracker based on class.

7

I’ve seen this but this would apply equally to all classes, right? I don’t want to degrade the tracking quality of other classes if I enable results from shadow tracking but maybe I can just pick certain classes from there?

8

Yes, the config parameters will apply to all classes. Why you want different config parameters for different class? Can you share some videos to show the issue?

9

Sorry, I’m not able to share the video since I don’t have the rights for it.

The reason I want different parameters is the exact same as stated initially. I have classes that I need to track, the parameters are tuned to those classes. Other classes I don’t need to track. Now if I tune the tracking parameters to those classes that I don’t need to track, obviously this will impact the classes that I need to track.

I have found a solution already, what I do now is to remove the objects that I don’t need for tracking before the tracking module and then after tracking I add them back in.

Is it possible for you to add a new feature to tracking? One where you can exclude certain classes from being tracked and just pass their raw detections downstream? I can imagine that this wouldn’t be too hard to add, since you already have a way to ignore ALL classes.

10

Glad to know you have a solution for the issue.

Thanks for you suggestion. Seems nvtracker need handle different class separately.

1 Like
11

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.