When new category of objects occur in the later frame, which are not prompted in the first frame, are there some good methods to tracking the new objects?

[RongkunYang]

No description provided.

[rentainhe]

Thank you very much for raising an excellent issue. We believe this is an engineering problem. We have attempted to implement a version that may require detection for each frame and then determine if it is the same object. However, the accuracy of this process cannot be guaranteed. If there is occlusion between some objects in specific frames, it is likely that the consistency of the ID cannot be ensured. But if it can be combined with some tracking models to ensure the consistency of the ID, it might still be possible to achieve. @RongkunYang

[ShuoShenDe]

As the previous answer, to track new objects that appear in later frames (which were not present in the initial frame), iteratively detecting objects in each new frame is necessary. When a new object is detected, you have two main approaches to handle it:

1. Restart the Tracking Model with Updated Object List: Stop the existing tracking session and start a new tracking instance that includes the newly detected objects. This approach is used in methods like grounded_sam2_tracking_demo_with_continuous_id, which allows continuous tracking by updating the object list dynamically.
2. Use a New Model Instance for Tracking the New Object: Another approach is to create a separate tracking model specifically for the new object, allowing the original tracking instance to continue independently. This method, similar to tracking_recursively, avoids interruption to the original tracking but is computationally heavy.

While neither of these methods is truly real-time, we continue working toward a real-time tracking solution for dynamically appearing objects.