Normalization free Siamese network for object tracking

Siamese-based trackers have received global recognition for target tracking. However, these trackers employ batch-normalized networks for feature extraction, which has been sensitive to batch size and hard to replicate on different hardware. Therefore, to meliorate tracking performance and effectively address this issue, the present work proposes a Normalization free Siamese (NfS) tracker by introducing normalization-free networks in target tracking. The developed NfS has been trained end-to-end with large-scale datasets such as COCO, TrackingNet, LaSOT, VID, DET, and GOT10k. Extensive experimentation has been carried out on six challenging benchmark datasets (OTB100, LaSOT, VOT2018, VOT2019, UAV123, and GOT10k), revealing that NfS ensures comparable performance with state-of-the-art (SOTA) trackers on most of the benchmarks. It pushed the performance bar by a minimum of 2.88% and 2.37% on UAV123 for both precision and success scores. Also, it overshadows the compared trackers by a significant minimum margin of 11.88% and 8.14% on the LaSOT for similar metrics, demonstrating the higher discrimination capability of the NfS tracker for both natural and aerial target tracking tasks.