Face Detection and Recognition in videos

Vaibhav Sharma's picture

GSOC project progress

Some more testing today, the accuracy improved after applying PCA to input features. Now the accuracy i am getting is 45% on IMFDB dataset which is much better than 15-18% on videos for 6 classes. Although one problem that i faced while testing due to increase in the number of testing vectors is the increase in time that the classifier takes to produce results. It takes approximately 10 minutes for one run of the classifier for testing of 230 features.

Vaibhav Sharma's picture

GSOC project progress

Today collected some results using IMFDB dataset and solved some of the errors that were caused by difference in decision function of different classifiers of scikit-learn library. Today's results were a bit better, the new method of preprocessing images and then applying PCA is giving better results.

Vaibhav Sharma's picture

GSOC project progress

Today worked on cross checking my code with one of the papers from where i was following the whole methodology. The feature extraction method the authors of the papers used was Local binary pattern but while using the same method i could not see any difference in my accuracy.I also worked on my MATLAB codes and added examples to them.

Vaibhav Sharma's picture

GSOC project progress

Today went through the paper on classification of IMFDB data set using semi-supervised learning. Using their methodology i tried feature extraction using "Local Binary Pattern". The results were almost same and it was hard to tell if the new method was better or worse. I would do more testing with it on my video to see if the results can get better.

Vaibhav Sharma's picture

GSOC project progress

Not something very concrete that i did today.But i spent sometime reading and researching about different ways to extract feature vectors from images. The previous algorithm 'Contourlet Transform' that i mentioned, i am still understanding its basics and techniques involved in it and would implement it as soon as i understand it. The new algorithm that i saw today was to consider polar coordinates of each pixel and then extract features from the radii and angle. I am still reading about it and have started working on extracting features through radii of pixels.

Vaibhav Sharma's picture

GSOC project progress

Not much work today. Just added the MAD example to the repository and made some changes in the GUI code to make two separate running examples of MAD and LDS.

Vaibhav Sharma's picture

GSOC project progress

Vaibhav Sharma's picture

GSOC project progress

Today finally I was able to produce a nice looking GUI to handle all the working of the code and provide more usability. The GUI is made using PyQT and i am using multi threading to run my code. There were still some issues with the GUI and code integration but i think for now i would ignore that stuff and focus on the main parts of my project again and begin working on accuracy part.

Vaibhav Sharma's picture

GSOC project progress

Today also spent most of the time on Testing and trying different things for better accuracy. In addition to that i spent hours struggling to make a GUI that can play openCV videos and also provide multiprocessing feature of running my code simultaneously. I could not make the Tkinter code work, the program used to get stuck in the middle and hang indefinitely. So i finally went on to use PyQT and exploring its features for a better GUI.

Vaibhav Sharma's picture

GSOC project progress

Today i was able to find some of the reasons for the low accuracy of the program. The scikit library i was using was actually giving very low accuracy on the data and it was able to classify only a few vectors with high confidence, because of which my classifier got only a few correctly labelled vectors. So today's findings were important as it showed that there is scope of higher accuracy if a better supervised classifier is used and that my classifier has better accuracy than it is giving now, it's just a matter of correctly labelled input vectors.

Syndicate content