using proper set of offset for calculating L_seg and L_angle losses

[nagitam]

Thanks for the very interesting work. At the end of section 4, you have mentioned Since the NMS block is not differentiable, the only way for the network to minimize L_seg and L_angle is to generate a proper set of offsets t for Equation 6.

Can you please elaborate more on this? Does this mean you are not using the t_i values generated by the attentional gnn? Do you propagate this loss through the score matrix generated by the optimal connection network?

[zorzi-s]

Hi @nagitam,

thank you for the interest on our work!

The offsets $ti$ are used to refine the vertex positions minimizing $L{seg}$ and $L_{angle}$. The final positions are given with equation (6): $\hat{p}_i = p_i + \gamma \cdot t_i$. Since the NMS layer, that produces the initial coordinates $p_i$, is not differentiable, the network can just learn to modify the positions through the offsets $t_i$. The refinement losses do not affect vertex detection map and NMS, and therefore, the coordinates $p_i$.

[nagitam]

Thank you @zorzi-s Your explanation makes sense 👍