Strange cross shape in quality prediction

[destroy314]

The above figure shows the object image and the result of the grasp quality prediction (q_img in code). It can be seen that the red cube becomes a slightly larger cross in the grasp quality prediction, resulting in a wrong grasp prediction (the red line above the red square). Does anyone know why this is or how to avoid it?

[destroy314]

Trained a model on cornell dataset, looks better now.