ybsong00 / crest-release Goto Github PK

CREST: Convolutional Residual Learning for Visual Tracking

MATLAB 33.98% Makefile 1.10% TeX 9.73% Python 5.99% CSS 0.36% JavaScript 0.05% HTML 0.47% C++ 13.78% Cuda 31.32% C 1.33% Shell 1.87% M 0.03%

crest-release's People

Contributors

Stargazers

Watchers

Forkers

whjzsy zhengzhugithub opencvfun statml zhangxujinsh benjamesbabala fireae ouya-bytes lyk125 zhu1teng tcyhx hxl1990 mrdoer collector-m prozyy wangyx1990 zgsxwsdxg hxkwindwizard wangxianliang azuredsky xiongduan yisongshen starstylesky rikiw dgqsyujian wolfworld6 kimsoohwan damax18 dlrookie pangye merlin2013 walkoncross yangjessie guangshengshi guoyaxiang wpfhtl abdelpakey xiuxiuzhang1995 chenxiangxie1996 yungu-imr liujiandu zjuxwg0401 alei089 afcarl sookim-ai luvios cyberspacefighter sysu-shey xrosliang wf1024966

crest-release's Issues

Error using vl_argparse when OTB test

using vl_argparse error when OTB test:
Expected either a param-value pair or a structure

about testing code

i find the code can not run in OTB100，because the size of the object is different，but in code，it designed resize the scales of 5 times bigger than target. can you give me a copy of test codes which can test on the OTB or VOT，thanks,

the false of using v1_nnconv

错误使用 vl_nnconv
The option name is not a string (argument number 5)

出错 dagnn.Conv/forward (line 11)
outputs{1} = vl_nnconv(...

出错 dagnn.Layer/forwardAdvanced (line 85)
outputs = obj.forward(inputs, {obj.net.params(par).value}) ;

出错 dagnn.DagNN/eval (line 91)
obj.layers(l).block.forwardAdvanced(obj.layers(l)) ;

出错 CREST_tracking (line 57)
net1.eval({'input',gpuArray(patch1)});

出错 Demo (line 32)
result=CREST_tracking(opts,varargin,config,display);

could you help me to solve this？

ICCV 2017 spotlight?!

The idea is apprently borrowed from the other one convolutional layer-based tracker CRT, but not cited.
And I can not really understand how the temporal residual learning in the proposed framework can be helpful for tracking. It is unreasonable!

Cant believe this is ICCV 2017 spotlight.

run Demo.m Error

outputs{1} = vl_nnconv(... inputs{1}, params{1}, params{2}, ... 'pad', obj.pad, ... 'stride', obj.stride, ... obj.opts{:}) ;

Error using vl_nnconv
The option name is not a string (argument number 5)

Error in dagnn.Conv/forward (line 12)
outputs{1} = vl_nnconv(...

Error in dagnn.Layer/forwardAdvanced (line 85)
outputs = obj.forward(inputs, {net.params(par).value}) ;

Error in dagnn.DagNN/eval (line 91)
obj.layers(l).block.forwardAdvanced(obj.layers(l)) ;

Error in CREST_tracking (line 57)
net1.eval({'input',gpuArray(patch1)});

Error in Demo (line 32)
result=CREST_tracking(opts,varargin,config,display);

Could you help me, how to solve it? I have successed compiled matconvnet

Error using vl_argparse

Dear author,

i have installed matconvnet successfully. When running Demo.m, i got some errors. Just showed below:
Error using vl_argparse (line 91)
Expected either a param-value pair or a structure.

Error in cnn_train_dag (line 26)
[opts, varargin] = vl_argparse(opts, varargin);

Error in CREST_tracking (line 117)
info = cnn_train_dag(net_online, imdb, input,getBatchWrapper(opts), ...

Error in Demo (line 32)
result=CREST_tracking(opts,varargin,config,display);

I think that there is something wrong with your input variable values in the function "cnn_train_dag", or "getBatchWrapper", but i can't find the solution about that error. Has anyone else had the same problem when running Demo.m?

Thank you for your time, i am looking forward to your reply.

About the network design?

@ybsong00 I noticed that the CREST only use 22 layers of VGG. By using all the convolution layers, will the accuracy increase? And the learning rate of the residual network is too slow. Does it mean that the parameter of the network need not to be changed substantially. Therefore, can the network be trained offline with large tracking benchmark, thus making the CREST faster without much online training?

run Demo.m Error---dimension doesn't agree

`Error using *
Inner matrix dimensions must agree.

Error in CREST_tracking (line 96)
feat_=feat_*coeff;

Error in Demo (line 32)
result=CREST_tracking(opts,varargin,config,display);

Error in run (line 86)
evalin('caller', [script ';']);`

Hello, here is the problem, feat_: 1369 x 512 single, and coeff: 1369 x 64 double, can you help me solve it?

Can you provide vot trax code?

@ybsong00 I am interested in your great job! Can you provide vot trax code so that I can test CREST in VOT2016 and VOT2017?

About the effect of residual layer ?

@ybsong00 @dynamicsong Can the residual layer really capture the appearance changes ? I think it only contributes to the training of base layer.

scale change problem?

Hi， @ybsong00 . I am interested in the CREST and tested it in OTB and VOT. I found that the performance drops when the scale of the target change rapidly compared to other trackers. Maybe the search ratio [1, 0.95, 1.05] is not a good solution. Do you have some advices to solve this problem? Thanks!