search

liboyin / img-classify

Household Object Recognition Challenge
GNU General Public License v2.0
0 stars 0 forks source link

Sakda's project timeline #8

Open liboyin opened 9 years ago

liboyin commented 9 years ago

Apr 17: Followed Sentdex Image Recognition. This gave real bad result, might need to project, rotate, translate images before testing.

sakda-mail commented 9 years ago

[Apr 21] Review: shackenberg/Minimal-Bag-of-Visual-Words-Image-Classifier

[Apr 22] Setup environment and libraries for Shackenberg code.

[Apr 24] Be able to run Shackenberg code on a Ubuntu machine.

[Apr 25] Re-organized the image file, some of file names are skipped, modify the code to read all image from 'dataset' folder, categorize and pick one for a tester. The code run properly but result is only 60% correct.

[Apr 26] Be able to fix path of libraries in order to run on 32-bit Windows machine.

[May 2] Come up with an idea using SiftMatch to classify images. The assumption is for any image in a same category should have a highest number of match points.

[May 3] A new model using SiftMatch is created. The accuracy is about 80-85%. This technique can predict very well on label of goods, but not for toys and fruits. Running the code on multi thread can improve speed and smoothing the images before sift can improve the accuracy a bit. (using kernel=-1 * ones(3)/9;). Thus, the accuracy is about 85-90%.

[May 9] BoW model is created on Matlab version. The accuracy is about 80%.