ORB and Object Detection

[yash-chowdhary]

I tried feature matching using ORB and SURF (mainly because getting SIFT descriptors was time-consuming). The results were horrible, to say the least. If I tried to extract the keypoints that matched, we'd end up getting a large portion of the image as waldo/wenda/wizard.

The other approach I tried is object detection - similar to lab 4. I split the templates for the 3 characters into training and testing sets (70-30). I created my vocab using the training images. Then, I extracted sift features of each template for each character using dsift, and created bags of sifts. Finally, I trained an SVM and recorded its accuracy.

[khooroko]

it would be good to leave comments or write somewhere which cells to run 😅 many of the cells can't be run due to various errors such as unexpected indent. Also the .DS_Store shouldn't be committed

[yash-chowdhary]

Added a few comments. The second cell is a mixture of a lot of methods and is really convoluted. It can be ignored for the time being.

[khooroko]

Would it help to train 3 separate vocab files for the 3 classes (waldo, wendy, wizard)? 🤔

[yash-chowdhary]

there was some sort of a mixup while committing my code locally. i've cleaned it up now

[yash-chowdhary]

Would it help to train 3 separate vocab files for the 3 classes (waldo, wendy, wizard)? 🤔

Regarding this, I'm not so sure about this.. Isn't your vocabulary supposed to be general, and not specific? This way, it can be more robust?
If we have 3 different vocabularies, we'd need 3 different classifiers instead of the one classifier that classifies 3 labels.

Either way, we might need to train our SVM with random images as well, so that pictures that are not any of Waldo, Wendo or Wizard don't end up getting classified as one of the 3.