zhang-pengyu/adrnet

code and results for 'Learning Adaptive Attribute-Driven Representation for Real-Time RGB-T Tracking'

adrnet's Issues

请问代码还要多久才能发布？

首先，你们的工作非常不错，对我的工作很有启发，因此我想去通过代码更深入地了解你们的工作。但是似乎代码公布没有一个确定的日期，希望能够尽快地发布，然后一览你们的工作。

多步训练的代码

您好！请问多步训练的代码会开源吗？

师兄，请问这个算法的速度大概是多少？

师兄，请问这个算法的速度大概是多少？

RT-MDNet权重

您好，我想问一下为什么你提供的RT-MDNet的权重有1G多大，而原本的RT-MDnet的权重只有大概16M

Run Demo Error

The following error is reported when the demo sequence is run:
UnicodeDecodeError: 'utf-8' codec can't decode bytes…………
Wish your reply!

RuntimeError: Error building extension '_prroi_pooling'

when I run python Run_test.py, I have encountered this problem, could you give me some ideas to solve it

SCANet

您好，请问SCANet.pth哪里可以下载，我看代码里用到了它的权重。

preciseroipooling的使用

作者大佬，我使用git clone下载的你的代码，我的pytorch是0.4的，没有找到./travis.sh这个文件

How to evaluate the results?

Hello,
Following the README file, I run the code and have got some results, i.e. the predicted bounding boxes. But, it seems not be mentioned that the evaluation protocal you use.
Q: Should I use the official RGB-T 234 evaluation tools? or other evaluation tools?
Thank you!

ImportError: Can not compile Precise RoI Pooling library?

when we run "python Run_test.py“，there is a problem: ImportError: Can not compile Precise RoI Pooling library? thanks

论文放到ａｒｘｉｖ了吗

这个是基于Siamese框架还是RT-MDNet框架呀

关于多步训练

我想请问下,在论文里的分步训练的第二步训练AENet时,论文里说要冻结主干.ADRB和FC,这里的FC层,是FC4.FC5.FC6吗,包不包括FC6啊,还有就是怎么最小化公式6的那个权重Ma,这边没怎么看明白,希望作者大大能看到,回复一下

no module named 'sample_generator'

How to multi-step train the model?

In the paper, this model adopts a multi-step training strategy, which consists of three steps. But I can't find how to use train_ADRNet.py for multi-step training? Could you give me some suggestions?

数据集下载

您好，能不能给个代码相关数据集的下载链接？

训练

学长，你好，请问train_ADRNet代码部分的SCANet.pth是做什么用的呀，我看网盘里面没有这个文件？

About CENet

Great work!
I have a question about your CENet mentioned in paper.
Why did you use method of summation and 5 fc layers to handle five residual features rather than method of concat and 1 fc layer?
It seems quite novel but hard to understand.
Wish your reply!

训练过程中如何得到iou的值

try:
total_miou = sum(total_iou)/len(total_iou)
except:
total_miou = 0.
print ("Mean Precision: %.3f Inter Loss: %.3f IoU: %.3f" % (prec.mean(), totalInterClassLoss.mean(),total_miou))
作者大佬，你在train_ADRNet.py 代码的196行异常捕捉这部分代码中，total_iou似乎没有定义，这样运行结果中的iou会总是等于0，关于iou的值该如何得到呢？

关于ADRNet 中各组件的消融实验

您好,我想请教一下
1.论文里的表4,B(RGB)和B(T)是怎么得到相应的数据集结果的,是把输入双流网络的图像都换成RGB图像或都是T图像吗?
2.B(RGBT)的成绩是由feat_RGBT 还是 residual_none得到的呀,我在实验时发现feat_RGBT的成绩会好一点,我就比较困惑那为啥还要有一个none分支,直接把none分支由feat_RGBT替换不行吗?论文里说是因为还有其他的属性都包含在none分支中,我想了解下是否还有其他的原因?
希望作者大大不吝赐教,感谢感谢.

about Evaluation Index

Hello! Thank you very much for your contribution to the field of RGBT tracking. After reading your paper and reproducing your code, I benefited a lot.
I would like to ask you a question about indicators. Are your indicators on GTOT and RGBT234 realized through the official toolbox? I see that the txt file of Guangfang requires 8-dimensional data, but the data you generated is 4-dimensional. If possible, can you provide your evaluation plan? Or your final results file for evaluation. I will be very grateful!

no "install.sh"?

SCANet.pth

您好，请问一下为什么在测试阶段没有使用bbox regression优化box？

Evaluation on RGBT234 toolkit

Hello, when I was evaluating the tracking results, the RGBT234 toolkit required each line of txt to have 8 values, but the results I reproduced only had 4 values in txt. I would like to ask how I should modify it.

关于多步训练的一些问题

作者大大您好，我在复现代码时发现，最后一步离线训练微调FC时，为什么FC微调之后性能反而下降了，请问这里面是有什么不同的参数设置或者技巧吗？微调FC时，所有的FC layers的学习率设置的都是一样的吗？