Relaxed Rotational Equivariance via $G$-Biases in Vision

The AAAI2025 accepted this paper! We will release the code in the next few days.

Note: Unlike the experiments reported in the paper, we simplified the classification header and re-conducted the experiments, resulting in more prominent results with fewer parameters. The new results are as follows:

Experiments on the CIFAR-100 / 10 dataset

Model Name	Group $G$	Equivariance	CIFAR-100 Acc.	#Param.	CIFAR-10 Acc.	#Param.
c2_sre_n	$C_2$	SRE	76.95	325748	-	314138
c2_rre_n	$C_2$	RRE	77.84	345140	-	333530
c4_sre_n	$C_4$	SRE	80.79	625012	-	-
c4_rre_n	$C_4$	RRE	82.48	663796	-	-
c6_sre_n	$C_6$	SRE	81.36	924276	-	-
c6_rre_n	$C_6$	RRE	83.20	982452	-	-
c8_sre_n	$C_8$	SRE	82.52	1223540	-	-
c8_rre_n	$C_8$	RRE	83.63	1301108	-	-

Model Name	Group $G$	CIFAR-100 Acc.	#Param.	CIFAR-10 Acc.	#Param.
c4_rre_n	$C_4$	82.48	663796	-	-
c4_rre_s	$C_4$	84.47	2524548	-	-
c4_rre_m	$C_4$	85.35	8061316	-	-

You can train the model using the following command:

python classification/train.py -model [Model Name] -dataset [Dataset] -batch_size 128

[Model Name]: c2_sre_n, c2_rre_n, c4_sre_n, c4_rre_n, c4_rre_s, c4_rre_m, c6_sre_n, c6_rre_n, c8_sre_n, c8_rre_n

[Dataset]: cifar10, cifar100

Example command:

cd classification
python train.py -model c4_rre_n -dataset cifar100 -batch_size 128