4

Most of the times I see people using the RGB channel for the training set. While it is good for facial detection but I have not found it as effective as I thought it would be! So, should I try converting the image into a different (possibly a combination of different color transitions) color space to train an object identifier?

I cannot afford to do find it out experimentally as there is a higher cost associated with that(AWS).

P.S. I am using dlib's dnn_mmod_ex.cpp!

PanNik
  • 1,564
  • 1
  • 8
  • 11
  • 1
    Some related links https://www.reddit.com/r/MachineLearning/comments/3chylb/is_there_an_advantage_to_encode_images_in_yuv/ http://users.ics.aalto.fi/perellm1/thesis.shtml – Thomas Wagenaar Jun 15 '17 at 16:23

2 Answers2

2

this paper might be interesting for you:

"Effect of image colourspace on performance of convolution neural networks" by K Sumanth Reddy; Upasna Singh; Prakash K Uttam. Published in: 2017 2nd IEEE International Conference on Recent Trends in Electronics, Information & Communication Technology (RTEICT).(https://ieeexplore.ieee.org/document/8256949)

The authors investigated the effect of different colorspaces (RGB, HSL, HSV, LUV, YUV) on the performance of a CNN (AlexNet) trained with CIFAR10 dataset. They found that the LUV colorspace is a good alternative to the widely used RGB colorspace, while the Network trained on YUV data showed the worst performance.

femonk
  • 65
  • 1
  • 10
  • 1
    The abstract: “We have observed that LUV colourspace is the best alternative to RGB colourspace to use with CNN models with **almost equal** performance on the test set of CIFAR10 dataset.” (emphasis mine.) Thus they found that RGB is the best color space. – Cris Luengo Oct 16 '18 at 13:04
  • 1
    That's right @CrisLuengo. However, they conclude _"Although RGB colour space is quite popular with CNN models, there any many other colourspaces like HLS, HSV, LUV, YUV etc. A comparison has been made between these colourspaces with respect to the performance of the CNN model on the test set of the CIFAR10 dataset and it is concluded that LUV colourspace is a best alternative to work with CNN model while YUV colourspace is the worst one. And to our knowledge this is the first time such an experimentation is done to understand the performance of CNN on different colourspaces."_ – femonk Aug 10 '19 at 10:33
1

After long search for answers and experimenting a lot with the color space in object recognition tasks, I've found out RGB is an optimal choice for the color space. I tried many color Spaces such as YCrCb, Lab etc. While they are certainly great when we are limited to computer vision problems like Lab is exposure invariant(upto some extent) but an image certainly lose data when we convert it to other colorspaces. I was using HOG as a feature descriptor and it didn't go well with changed color spaces. The gradients were misfiring and there was no consistency. However in computer vision, where I am dealing with defect segmentation in potatoes, other color spaces such as YCrCb proved to be a great feature descriptor for defect segmentation via clustering algorithms. Do correct me if you find some thing incorrect in my observations.

PanNik
  • 1,564
  • 1
  • 8
  • 11