Is it useful to have only one convolutional layer in the feature alignment module ？

[houjun0322]

I found that the feature alignment module only has a deformable convolution and a 3x3 convolution, does the presence or absence of this module have a big impact on the results?

[houjun0322]

In addiction, I found in the code that in the weighted average of the final regular classification score p_r and feature alignment score p_o, the p_r accounted for 0.94 and the p_o was only 0.06. https://github.com/researchmm/TracKit/blob/e351e5bff8071aa14f333d5975a8f408a3e264c6/lib/tracker/ocean.py#L274