Enhancing Transformer Tracking Using NF-ResNet and ResNeXT Backbones

Residual networks play a foremost role in the domain of tracking, specifically in the extraction of features. The residual networks are using a simple technique of skipping connections to overcome the problem of vanishing gradient. Also, it is using batch normalization after each layer of convolution to accelerate the training by reducing the dependency over the parameters. Though, we have tried to solve these issues by introducing ResNeXt and NF-ResNet one by one into the TransT tracker as feature extractors in the place of ResNet. Finally, we present a ResNeXt-based TransT tracker (named as TransT_NeXt) and NF-Net-based TransT tracker (named as TransT_NFNet). We have evaluated proposed trackers on three large-scale benchmark datasets and four small-scale benchmark datasets but unfortunately, our trackers haven’t performed well. At last, we have discussed about the future work on the basis of this study.