[Retracted] Relieving the Incompatibility of Network Representation and Classification for Long-Tailed Data Distribution

Table 7

Long-tailed classification results on Places-LT, starting from an ImageNet pretrained ResNet-152.

Model	Many-shot	Medium-shot	Few-shot	Acc
ResNet-152	>100	≤100 and >20	≤20
Plain model	45.5	27.8	8.5	30.2
Many-shot model	46.4	—	—	—
Lifted loss [41]	41.1	34.8	22.4	34.6
Focal loss [13]	41.1	35.4	24	35.2
Range loss [14]	41.1	35.4	23.2	35.1
FsLwf [42]	43.9	29.9	29.5	34.9
OLTR [10]	42.2	38.1	17.8	35.3
BALMS [43]	41.2	39.8	31.6	38.7
Decouple (cRT) [12]	41.6	39.4	29.2	38.1
Decouple (-normalized) [12]	37.8	40.7	31.8	37.9
Ours with OLTR	43.8	37.8	17.5	37.5
Ours with decouple	41.5	40.9	32.2	38.7

Results directly copied from Ref. [10]. Results reproduced with author's code.