Transformer with Decoupled Query Augmentations for End-to-End Multi-Object Tracking

What is it about?

In many real-world situations, such as surveillance and autonomous driving, computers need to detect and follow multiple moving objects at the same time. While recent AI models based on Transformers can handle both tasks in an end-to-end manner, training them jointly can cause the tasks to interfere with each other. This work introduces a new AI model that reduces this conflict between detecting new objects and tracking existing ones, while still keeping the model fully end-to-end. We achieve this by separating key steps in the model’s processing pipeline and designing each part specifically for either detection or tracking. As a result, the model becomes better at both identifying new targets and consistently following those it has already seen.

Featured Image

Photo by Korhan Eser on Unsplash

Photo by Korhan Eser on Unsplash