notesum.ai

Published at November 14

SAG-ViT: A Scale-Aware, High-Fidelity Patching Approach with Graph Attention for Vision Transformers

cs.CV

cs.AI

cs.LG

68T07

I.2.10

Released Date: November 14, 2024

Authors: Shravan Venkatraman¹, Jaskaran Singh Walia¹, Joe Dhanith P R¹

Aff.: ¹Vellore Institute of Technology, Chennai, India

Arxiv: http://arxiv.org/abs/2411.09420v1

Refer to caption

Backbone	CIFAR-10	GTSRB	NCT-CRC-HE-100K	NWPU-RESISC45	PlantVillage
DenseNet201	0.5427	0.9862	0.9214	0.4493	0.8725
Vgg16	0.5345	0.8180	0.8234	0.4114	0.7064
Vgg19	0.5307	0.7551	0.8178	0.3844	0.6811
DenseNet121	0.5290	0.9813	0.9247	0.4381	0.8321
AlexNet	0.6126	0.9059	0.8743	0.4397	0.7684
Inception	0.7734	0.8934	0.8707	0.8707	0.8216
ResNet	0.9172	0.9134	0.9478	0.9103	0.8905
MobileNet	0.9169	0.3006	0.4965	0.1667	0.2213
ViT - S	0.8465	0.8542	0.8234	0.6116	0.8654
ViT - L	0.8637	0.8613	0.8345	0.8358	0.8842
MNASNet1_0	0.1032	0.0024	0.0212	0.0011	0.0049
ShuffleNet_V2_x1_0	0.3523	0.4244	0.4598	0.1808	0.3190
SqueezeNet1_0	0.4328	0.8392	0.7843	0.3913	0.6638
GoogLeNet	0.4954	0.9455	0.8631	0.3720	0.7726
Proposed	0.9574	0.9958	0.9861	0.9549	0.9772