Open NomiMalik0207 opened 5 years ago
Yes I append the landmarks with the hog features and the flatten image, then I feed everything to the SVM. Since it's supervised learning, the SVM expects also the labels so that it learn to classify each data to the correct label. But the labels are not appended with the data
But from you model Picture it seems you are feeding your model only with image, and sending landmarks as a different layer, Can you explain a little bit more about feeding landmarks? Thank you for your time
In this svm version I concatenate the two, but in the other repository using cnn indeed I feed the images alones in a 2d convolutional layer, and I feed the landmarks (pixels positions) in another layer. In some cases I don't feed the landmarks alone, but I append the hog features, thoses were just tests I was trying to see how feeding an image alone, and feeding additional data which are not raw pixels can affect the training
Got it. But i have tried a CNN model, and including landmarks did not effect the accuracy.. :(
What was the model you used?
VGG
I have a hypothesis but not sure: may be the landmarks are useful only on small architecture, like the 5-layers model I used, but in the case of bigger models having dozen of layers the parameters space is huge and the network can optimize the solution better, without the need of extra features such as landmarks of HOG In other words, in bigger CNN models, the landmarks provided may be just redundant information
I want to know about how you ae feeding your network with landmarks? are you appending landmarks on image or just using as a label?