MixFormerV2 icon indicating copy to clipboard operation
MixFormerV2 copied to clipboard

[NeurIPS 2023] MixFormerV2: Efficient Fully Transformer Tracking

Results 25 MixFormerV2 issues
Sort by recently updated
recently updated
newest added

how to judge whether targets disappear in Mixformerv2. Currently,trying to use confidence scores to judge, sometimes the results are Untrustable. Thanks

Thank you for your impressive paper. I noticed that there are two pre-trained models available on the Google Drive, 'base' and 'small'. I'm unsure if the 'base' pre-trained model corresponds...

首先,感谢您优秀的工作,由于我是一名目标跟踪的小白,也是乞丐科研工作者(配置32G显存),训练后自测发现,序列模型的训练效果和训练时的batchsize或者序列图片多少成正比,我想咨询一下是不是有这方面的问题,这很关键,可能导致序列模型无法复现的问题,再次感谢您的关注,欢迎大家加入讨论,给指导一条明路

Hello,首先谢谢你们提供的算法,很棒的工作。我目前在研究算法的部署。 关于各个数据集上的指标,我有一个小问题:AUC和Norm Precision,哪一个更适合作为首要的评价指标。 我个人的理解是,Norm Precision应该更能表现跟踪的准确度,而AUC则更反应跟踪算法的鲁棒性?

I have been trying to see the process of training MixformerV2-S (the small one). But I think the necessary files are not included, like YAML files. Can you please share...

Hi,there 按照`tracking/test_mixformer.sh`脚本的指令在got10k和lasot上测试的跟踪结果都是无效。 加载的是mixformer2_base.pth.tar权重,网络选择的是mixformer2_vit_online,参数选择288_depth8_score的配置。在A100单卡上面进行推断,请问是使用方法不对吗?

首先,十分感谢作者的开源! 我在看代码时对于如何由(coord_l, coord_t, coord_r, coord_b)得到最终的预测结果看不懂,不知道为什么要与crop下来的search图像的边的长度相乘。也看不懂map_box_back和clip_box是在做什么。我看论文中说是“对四个坐标的概率分布的回归”,实在看不懂是什么定位策略。请作者能帮我解答。感谢! def map_box_back(self, pred_box: list, resize_factor: float): cx_prev, cy_prev = self.state[0] + 0.5 * self.state[2], self.state[1] + 0.5 * self.state[3] cx, cy, w, h = pred_box...

MixFormerV2和MixFormer在模型结构上有哪些相同和不同呢?