For the training of ID
Sisyphus-ZPJ opened this issue · comments
Hello, I am confused that the performance of ID in the paper is inconsistent.
In Table I, the metrics of EPE for ID (4 iters, 1/4 resolution) and ID (4 iters, 1/8 resolution) are 0.72 and 0.77. Based on the checkpoints you have released, these results are correct indeed.
However, in the ablation experiments, the results changed.
In Table IV, the metric of EPE for ID (4 iters) is only 0.88.
And in Table VI, the metric of EPE for ID (1/8 resolution) is only 0.79.
I would like to ask why the performance of the model in Table I is significantly higher than that in Table IV and Table VI. Compared to the training configuration described in Section IV A, are there any changes？
This is because, as we mentioned in the title of Table IV, we used a smaller model (with less feature channels for the encoder and convgru) to do the ablation studies.
Sorry, I didn't notice it before. Thank you for your help.