Table 3.

Loss $L$ on combined set (test set or training set) of finally trained models for compared methods

Methods	$L\downarrow$ (TeWoR)a	$L\downarrow$ (TeWR)a	$L\downarrow$ (TrWoR)a	$L\downarrow$ (TrWR)a
Fine-tune	5.299	5.257	5.298	5.256
IID-Offline	3.519	3.571	3.521	3.572
EWC	6.986	7.071	6.984	7.074
SI	4.899	4.938	4.897	4.934
BufferSize$\left|\mathcal{M}\right|=\mathbf{1}\mathit{K}$
ER	11.267	10.982	11.283	10.991
ER-RM	5.186	5.195	5.187	5.192
NI-WLb	5.172	5.061	5.172	5.060
NI-WL-RDb	4.356	4.391	4.358	4.391
NI-WL-RMb	4.766	4.830	4.769	4.831
NI-WL-RM-KDb	4.631	4.782	4.632	4.776
BufferSize$\left|\mathcal{M}\right|=\mathbf{5}\mathit{K}$
ER	4.182	4.188	4.182	4.188
ER-RM	4.596	4.759	4.599	4.760
NI-WL	4.740	4.790	4.737	4.792
NI-WL-RD	4.172	4.386	4.170	4.388
NI-WL-RM	4.400	4.319	4.402	4.320
NI-WL-RM-KD	4.162	4.200	4.168	4.197
BufferSize$\left|\mathcal{M}\right|=\mathbf{10}\mathit{K}$
ER	3.754	3.799	3.755	3.800
ER-RM	4.616	4.707	4.614	4.704
NI-WL	4.607	4.513	4.609	4.514
NI-WL-RD	4.542	4.510	4.540	4.508
NI-WL-RM	4.182	4.182	4.180	4.180
NI-WL-RM-KD	4.028	4.045	4.029	4.047

Bold underline, italic underline, and underline font highlights the first place, second place, and third place with same BufferSize, respectively.

^aTeWoR is short for test set without rotation, TeWR is short for test set with rotation, while TrWoR is short for training set without rotation, and TrWR is short for training set with rotation.

^bNI-WL, NI-WL-RD, NI-WL-RM, and NI-WL-RM-KD are four variants of our proposed method.