>> result = chnsPyramid <div class="snippet

chnsCompute ignores isInit,about elucideye/acf

headupinclouds commented on July 16, 2024

https://github.com/pdollar/toolbox/blob/1a3c9869033548abb0c7a3c2aa6a7902c36f39c2/channels/chnsCompute.m#L119-L136

% get default parameters pChns
if(nargin==2), pChns=varargin{1}; else pChns=[]; end
if( ~isfield(pChns,'complete') || pChns.complete~=1 || isempty(I) )
  p=struct('enabled',{},'name',{},'hFunc',{},'pFunc',{},'padWith',{});
  pChns = getPrmDflt(varargin,{'shrink',4,'pColor',{},'pGradMag',{},...
    'pGradHist',{},'pCustom',p,'complete',1},1);
  pChns.pColor = getPrmDflt( pChns.pColor, {'enabled',1,...
    'smooth',1, 'colorSpace','luv'}, 1 );
  pChns.pGradMag = getPrmDflt( pChns.pGradMag, {'enabled',1,...
    'colorChn',0,'normRad',5,'normConst',.005,'full',0}, 1 );
  pChns.pGradHist = getPrmDflt( pChns.pGradHist, {'enabled',1,...
    'binSize',[],'nOrients',6,'softBin',0,'useHog',0,'clipHog',.2}, 1 );
  nc=length(pChns.pCustom); pc=cell(1,nc);
  for i=1:nc, pc{i} = getPrmDflt( pChns.pCustom(i), {'enabled',1,...
      'name','REQ','hFunc','REQ','pFunc',{},'padWith',0}, 1 ); end
  if( nc>0 ), pChns.pCustom=[pc{:}]; end
end
if(nargin==0), chns=pChns; return; end

I believe isInit should be the C++ equivalent of the matlab condition that trigger the final line above if(nargin==0), chns=pChns; return; end.

That's probably what should happen here:

acf/src/lib/acf/chnsCompute.cpp

Lines 182 to 183 in 17b8da4

    
           //std::cout << pChns << std::endl; 
        
           // TODO

    if (!pChnsIn.complete.has || (pChnsIn.complete.get() != 1) || IIn.empty())
    {
        // Create default structures and merge:
        {
            // top level
            Options::Pyramid::Chns dfs;
            dfs.shrink = { "shrink", 4 };
            dfs.complete = { "complete", 1 };
            pChns.merge(dfs, 1);
        }

         // <SNIP>

        //std::cout << pChns << std::endl;
        // TODO
    }

Ah. TODO :)

This lib was created as part of an effort to get a few trained models running on a phone, so only some subset of the functionality has been tested. I'm not too surprised there are a number of use case gaps. Thanks for reporting. If you have a fix, feel free to send it. Otherwise, I can take a look later. We can add some GTests to cover some of this behavior as the issues come up.

The tests are here:

acf/src/lib/acf/ut/test-acf.cpp

Lines 382 to 400 in 17b8da4

    
           TEST_F(ACFTest, ACFChannelsCPU) 
        
           { 
        
               auto detector = getDetector(); 
        
               ASSERT_NE(detector, nullptr); 
        
               MatP Ich; 
        
               detector->setIsTranspose(true); 
        
               detector->computeChannels(m_IpT, Ich); 
        
           #if ACF_TEST_DISPLAY_OUTPUT 
        
               cv::Mat canvas = Ich.base().t(); 
        
               WaitKey waitKey; 
        
               cv::imshow("acf_channel_cpu", canvas); 
        
           #endif 
        
               // TODO: comparison for channels: 
        
               // load cereal pba cv::Mat, compare precision, etc 
        
               ASSERT_EQ(Ich.base().empty(), false); 
        
           }

TEST_F(ACFTest, ACFchnsCompute_isInit)
{
    // some test for expected default values with isInit here
}

from acf.

JN-Jones commented on July 16, 2024

Don't have any proper fix as I'm handling the options on my own for now. There's a similar construct in chnsPyramid though which seems to work:
https://github.com/elucideye/acf/blob/master/src/lib/acf/chnsPyramid.cpp#L179-L183

from acf.

headupinclouds commented on July 16, 2024

To be true to the reference implementation I suppose we should aim for the following:

>> result = chnsCompute

result = 

       shrink: 4
       pColor: [1x1 struct]
     pGradMag: [1x1 struct]
    pGradHist: [1x1 struct]
      pCustom: [0x0 struct]
     complete: 1

>> result.pColor

ans = 

       enabled: 1
        smooth: 1
    colorSpace: 'luv'

>> result.pGradMag

ans = 

      enabled: 1
     colorChn: 0
      normRad: 5
    normConst: 0.0050
         full: 0

>> result.pGradHist

ans = 

     enabled: 1
     binSize: []
    nOrients: 6
     softBin: 0
      useHog: 0
     clipHog: 0.2000

>> result.pCustom is empty

>> result.complete

ans =

     1

from acf.

headupinclouds commented on July 16, 2024

>> result = chnsPyramid

result = 
       pChns: [1x1 struct]
     nPerOct: 8
      nOctUp: 0
     nApprox: 7
     lambdas: []
         pad: [0 0]
       minDs: [16 16]
      smooth: 1
      concat: 1
    complete: 1

w/ >> result.pChns same as above

from acf.

headupinclouds commented on July 16, 2024

This should be consistent with the Matlab toolbox code now (after PR #40). I've added a couple tests to ensure that the default parameter retrieval is working and returning expected values. Closing ...

from acf.

chnsCompute ignores isInit about acf HOT 5 CLOSED

Comments (5)

Related Issues (20)

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

	TEST_F(ACFTest, ACFChannelsCPU)
	{
	auto detector = getDetector();
	ASSERT_NE(detector, nullptr);

	MatP Ich;
	detector->setIsTranspose(true);
	detector->computeChannels(m_IpT, Ich);

	#if ACF_TEST_DISPLAY_OUTPUT
	cv::Mat canvas = Ich.base().t();
	WaitKey waitKey;
	cv::imshow("acf_channel_cpu", canvas);
	#endif

	// TODO: comparison for channels:
	// load cereal pba cv::Mat, compare precision, etc
	ASSERT_EQ(Ich.base().empty(), false);
	}