dyhBUPT/iKUN

Suggestions for supplementing experimental data / 补充实验的建议

tylzh97 opened this issue · 5 comments

The method of using insertable Knowledge Unification Network for RMOT tasks is novel. The KUM module can extract visual features well, thereby improving the recognition accuracy of the target. The experimental part provides a new Refer-Dance dataset based on the DanceTrack dataset, which proves that ikun indeed has a good effect on the RMOT task. However, according to my observation, the target size in the dataset is relatively balanced. The article does not prove the tracking ability of ikun on different scales and different targets.
Can some experiments be added, such as tracking a picture of a person dancing and playing basketball at the same time, and tracking the positions of the basketball and the chicken person at the same time?
For example:
fbe1f235e16f9781e93f3ef8bd4dd88d

使用ikun进行RMOT任务的方法令人眼前一亮, KUM模块能够很好的提取视觉特征, 从而提高目标的识别精度. 实验部分在DanceTrack数据集的基础上, 提供了新的Refer-Dance数据集, 证明了在RMOT任务上ikun确实有很好的效果. 然而据我的观察, 数据集中的目标大小较为均衡, 文章并没有证明ikun在不同尺度上不同目标的追踪能力.
能否补充一些实验, 例如追踪一个人一边跳舞一边打篮球的动作, 同时追踪篮球与人的位置呢?

a good idea

wsw321 commented

hahaha

great idea makes me want to sing <Chicken you're so beautiful>, love from India ♥

+1

附议~

我还是被骗了