alexadlu / manet

Multi-Adapter RGBT Tracking implementation on Pytorch

MANet

RGBT234 dataset

链接：https://pan.baidu.com/s/1weaiBh0_yH2BQni5eTxHgg 提取码：qvsq

RGBT210 dataset

链接：https://pan.baidu.com/s/1FClmX0SH3WarcczkEQbmwA 提取码：ps8j

GTOT dataset

链接：https://pan.baidu.com/s/1zaR6aXh9PVQs063Q_b9zQg 提取码：ajma

RGBT234 toolkit

链接：https://pan.baidu.com/s/1UksOGtD2yl6k8mtB-Wr39A 提取码：4f68

RGBT210 toolkit

链接：https://pan.baidu.com/s/1KHMlbhu5R29CJvundGL4Sw 提取码：8wtc

GTOT toolkit

链接：https://pan.baidu.com/s/1iVVAXS4LZLvoQSGQnz7ROw 提取码：d53m

MANet result

Here, we have only uploaded the result file of the paper (PR_0.777 SR_0.539 on RGBT234, PR_0.894 SR_0.724 on GTOT.)

This code is an updated version, simplified from the one we submitted for the VOT2019-RGBT challenge.

Consequently, there are some differences compared to MANet's paper.

Prerequisites

CPU: Intel(R) Core(TM) i7-7700K CPU @ 3.75GHz GPU: NVIDIA GTX1080（8GB） Ubuntu 16.04

• python2.7
• pytorch == 0.3.1
• numpy
• PIL
• some others library functions

Pretrained model for MANet

In our tracker, we use MDNet as our backbone and extend it to a multi-modal tracker.

We use imagenet-vgg-m.mat as our pretrain model.

Train

You can choose either a two stage training or end2end training

two stage train:

• Stage1: Use the RGBT dataset to train whole network, then save the final model;
• Stage2: Load only the parameters of GA from the model saved in Stage1, and use the same RGBT dataset to train MA and IA while keeping GA fixed.

end2end train:

• The training method here is the same as with MDNet

Pretrain model :https://drive.google.com/open?id=1aO6LhOTxmpd7o_JXPLPjL3LsrQ5oqbl7

Run tracker

In the tracking/run_tracker.py file, you need to change the dataset path and save the result file directory. In the tracking/options.py file, you need to set model file path and set learning rate depend on annotation. For the testing and training stages, update the 'modules/MANet3x1x1_IC.py' file depending on the annotation.

Tracking model:https://drive.google.com/open?id=1Png508G4kQPI6HNewKQ4cfS36CvoSFSN

Result

manet's Issues

How to calculate Robustness in VOT2019-RGBT?

Hi, thanks for your work. I have a problem about calculate Robustness in VOT2019-RGBT, I used the official VOT toolkit but can not find how to get the result of Robustness in your paper, could you give me some suggestions? Thanks for your help.

vot-toolkit

你好，请问能否请教您一些关于vot测试的问题，谢谢

input image size during the training process

Are there any image size requirements during the training phase?

百度云？

Paper

你好，我的本科畢業設計是與rgbt跟蹤相關的，想拜讀一下您的代碼論文，方便的話能否給我一份論文？郵箱[email protected]

Result of MANet311-2IC model

Hello, I use the MANet311-2IC model you provide on RGBT-234 dataset, but the PR reslut is only 0.753 and SR is only 0.488, which is much worse than the 0.784 and 0.547 you provided. Thank you very much.

GTOT toolkit

Hello, I am using GTOT toolkit to evaluate the performance of model on different attributes. But I got the different results from your paper, canyou tell me how can I fix this? Thank you very much!
Below is the performance of different models on Thermal Crossover attribute:

How to test on VOT_RGBT2019?

Hello, I'm interested in your work. And I want to know how to test your code on VOT_RGBT2019. How to read the visible and ir data in the VOT_RGBT2019 running script?

how to use your code in three steps

Hi, I am newer in tracking.
I read your paper, and I find there are three steps to train your model.
Could you give some suggestions on how to use your code in three steps?
And what is the difference between training and tracking?
Waiting for your reply.

VOT result

Thanks for your great job,

Could you provide the results of VOT dataset, The result file is still not available on VOT official website.
Thanks for your help