LaTOT
收藏资源简介:
LaTOT是由安徽大学创建的大规模视频数据集,专注于微小目标跟踪。该数据集包含434个视频序列,总计超过21.7万帧,每帧都经过高质量边界框标注。数据集考虑了12种挑战属性,以覆盖广泛的视角和场景复杂性,旨在促进基于属性的性能分析。LaTOT不仅推动了微小目标跟踪的研究与发展,还提供了一个强大的基准方法,即多级知识蒸馏网络(MKDNet),该网络通过统一框架内的三级知识蒸馏有效增强了特征表示、区分和定位能力。数据集广泛应用于评估和验证MKDNet的优越性和有效性,为解决实际应用中的微小目标跟踪问题提供了重要平台。
LaTOT is a large-scale video dataset focused on tiny object tracking, developed by Anhui University. This dataset contains 434 video sequences, totaling over 217,000 frames, with each frame annotated with high-quality bounding boxes. It incorporates 12 challenging attributes to cover a wide range of viewpoints and scene complexities, aiming to promote attribute-based performance analysis. Not only does LaTOT advance the research and development of tiny object tracking, but it also provides a robust baseline method, namely Multi-Level Knowledge Distillation Network (MKDNet). This network effectively enhances feature representation, discrimination and localization capabilities through three-level knowledge distillation within a unified framework. The dataset is widely used to evaluate and verify the superiority and effectiveness of MKDNet, providing an important platform for solving tiny object tracking problems in real-world applications.

- 1Tiny Object Tracking: A Large-scale Dataset and A Baseline安徽大学 · 2022年



