About ablation study in Table 8 (w/o PCL-SD) #4
Description
Regarding Table 8 in the ablation study, I have two questions:
1.When "w/o PCL-SD" is indicated, which loss function was actually used—the CSD loss from A2S-v2 or the ADB loss from A2S-v1? The table only specifies "w/o PCL-SD," but the specific underlying loss function is not clearly stated.
2.According to the results in Table 8, adding PCL-SD only leads to a marginal improvement in pseudo-label quality (e.g., on RGB: Fβ increases from 0.895 to 0.896, M remains at 0.044; on NLPR: Fβ improves from 0.757 to 0.770, M decreases from 0.050 to 0.047). Why is the enhancement so limited? Could this be because the pseudo-labels have already reached a relatively high quality, leaving little room for further improvement, or is there another reason?