shipubupt/nips2018

Tracking

This is the implementation of our DAT paper "Deep Attentive Tracking via Reciprocative Learning ".
The project page can be found here: https://ybsong00.github.io/nips18_tracking/index.

The pipeline is built upon the py-MDNet tracker for your reference: https://github.com/HyeonseobNam/py-MDNet.
Note that our DAT tracker does not require offline training using tracking sequences.

Prerequisites

• GPU: NVIDIA GeForce GTX 1080 Ti
• CUDA 8.0.61
• python 2.7.14
• PyTorch 0.2.0_3 and its dependencies

Note

If you use our code based on a high-level version of PyTorch for other tasks, please ensure the "retain_graph=True, create_graph=True" in the backward function. Otherwise, the attention map cannot be used to update the parameters. Thank @Lu Zhou for checking the bug out.

Usage

1. Download VGG-M (matconvnet model) and save as "DAT/models/imagenet-vgg-m.mat"
2. cd DAT/tracking
   python demo.py

If you find the code useful, please cite both DAT and MDNet:

@inproceedings{nam-cvpr16-MDNET,
    author    = {Nam, Hyeonseob and Han, Bohyung}, 
    title     = {Learning multi-domain convolutional neural networks for visual tracking}, 
    booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},    
    pages     = {4293--4302},
    year      = {2016},
}

@inproceedings{shi-nips18-DAT,
    author = {Pu, Shi and Song, Yibing and Ma, Chao and Zhang, Honggang and Yang, Ming-Hsuan},
    title = {Deep Attentive Tracking via Reciprocative Learning},
    booktitle = {Neural Information Processing Systems},
    year = {2018},
  }

nips2018's Issues

training code

could you offer the training code?
thanks

[email protected]

请问
numpy pillow scipy matplotlib 等的版本是什么

Attentive

Attentive is the saliency?

what is the img_list in the function forwordgradients?

why it set 10 iterations in online update?

why it set 10 iterations in online update?

对于更高版本的 PyTorch，您的意思是所有的 backward function，都需要设置 "retain_graph=True, create_graph=True" ？

“If you use our code based on a high-level version of PyTorch for other tasks, please ensure the "retain_graph=True, create_graph=True" in the backward function. ”

对于更高版本的 PyTorch，您的意思是所有的 backward function，都需要这么设置吗？

where is the backward function? I cannot find the retain_graph=True, create_graph=True.

where is the backward function?
I cannot find the retain_graph=True, create_graph=True.

ImageNet

whether use ImageNet ?

Can you provide detailed instruction on how to reproduce the results in your paper?

@shipubupt , can you provide detailed instruction on how to reproduce the results in your paper? In this condition, it will facilitate the work of other researchers. Thanks a lot!

RuntimeError: cuda runtime error (2) : out of memory at /pytorch/torch/lib/THC/generic/THCStorage.cu:58

ubuntu16.04 ++940mx(2gb)+cuda8.0+py27+pytorch0.3.0post4 .I have changed batch_test=1 in option.py .but it is always out of memory,how to solve it?

这代码风格真的超烂，能不能重新写写？

With reciprocative learning , how do you draw the picture of the Visualization of attention maps on the Car4 sequence [50]?

With reciprocative learning , how do you draw the picture of the Visualization of attention maps on the Car4 sequence [50]?

thanks