An question on memory bank update with noise about ctvis HOT 1 OPEN

Hi @tomato18463 ,

Thanks for your attention on our work.

The reason we use noise is based on the former (Is this mechanism designed only to help the model recover when wrong id assignment happens in a specific frame (i.e. the model can reindentify the correct instance in the next few frames)). In our early experiments, we found that CTVIS is prone to ID switches under certain circumstances, such as when the target disappears, is obscured, or when the segmentation quality is poor, making it difficult for the model to recover the trajectory in subsequent frames. Therefore, we intuitively incorporated this process (update the memory bank with noise) during training to make the model more robust.

from ctvis.