Potential optimization of our model.

[yangtcai]

The paper DN-DETR presents a novel denoising training method to speed up DETR training, and they said that the bipartite graph matching algorithm we choose is unstable, and it may cause inconsistent optimization goals. So I think maybe we can adopt this idea into our model.

[williamstark01]

Is this model more or less complex than the original DETR?

[yangtcai]

That model is more complex than DETR, it optimizes the Hungarian match.

[williamstark01]

If you fully understand the algorithm they use and the math behind it and you think it might improve the performance of our network, it may be a good option to pursue.

That said, in my view anything that increases the complexity of the current model should be delayed until after we have a more robust training resulting to models with relatively good performance.