Hello, from the paper there's an instance interaction module that performs self attention among all instance features. The ablation also includes the different results with or without this module. But in the code I find there's no attention operation, in the each_stage_refine function, all the features are feed into a feed forwrard head to predict final boxes.