Go to ICLR 2025 Conference homepageClosed-loop Scaling Up for Visual Object Tracking
Thanks to the principles of the scaling law, current neural networks have experienced remarkable performance improvements. While much of the existing research has concentrated on upstream pretraining, the application of the scaling law to downstream vision tasks remains underexplored. Understanding the scaling law in downstream tasks can aid in the design of more effective models and training strategies. Thus, in this work, we aim to investigate the application of the scaling law to downstream vision tasks. Firstly, we explore the impact of three key factors of scaling law: training data volume, model size, and input resolution. We empirically verify that increasing each of these factors can lead to performance enhancements. Secondly, to address naive training's optimization challenges and lack of iterative refinement, we introduce DT-Training which leverages small teacher transfer and dual-branch alignment to further exploit model potential. Thirdly, building on DT-Training, we propose a closed-loop scaling strategy to incrementally scale the model step-by-step. Finally, our scaled model exhibits strong ability and outperforms existing counterparts across diverse test benchmarks. Extensive experiments also reveal the robust transfer ability of our model. Moreover, we validate the generalizability of the scaling law and our proposed DT-Training on other downstream vision tasks, reinforcing the broader applicability of our approach. We hope that our findings can deepen the understanding of the scaling law in downstream tasks and foster future developments on downstream tasks.