A General Data Augmentation Strategy for Siamese Object Tracking
Chaolin Pan, Jun Chu, Kai Huang, Lu Leng, Jun Miao · 2022 IEEE International Conference on Multimedia and Expo (ICME) · 2022
Siamese-based ResNet-driven trackers have made great success in recent years. However, the less discriminative features inhibit the improvements of these trackers. In this work, we discover that a possible reason is the emergence of the gridding artifacts in the deep layers of ResNet-driven trackers. We naively remove the deep layers with gridding artifacts, which reduces parameters by nearly 69% and improves the tracking performance of small objects. Then, we design a parameter-free feature superposition module by adding different samples into the first half samples in each mini-batch iteration in embedding space to increase the semantic information. Further, we introduce an auxiliary loss to reduce the learning difficulty. Finally, the generalization of the feature superposition module and the auxiliary loss are verified by ablation studies. Experiments on challenging benchmarks, including OTB2015, VOT2019, UAV123, LaSOT, and TrackingNet, demonstrate that the proposed method outperforms many SOTA trackers and achieves leading performance.