How to Label the image dataset?

asked 2019-01-19 09:09:27 -0500

Sky31 gravatar image

Hello folks,

I'm working on BOVW. I have some images lets say the car, bike etc. so my question is: How to give the label to the dataset.

edit retag flag offensive close merge delete


what are you trying to achieve ? what is the purpose of your program ?

are you trying to do a classification ? (then you need labelled samples) or a content based retrieval thing ? (then you don't)

creating a bagofwords dictionary itself does not require any labels (just a good ton of features), labels are only nessecary for a later ml based classification

berak gravatar imageberak ( 2019-01-19 09:45:19 -0500 )edit

yes, I'm working on the classification task. I need to classify the ID documents in the final project. currently, I'm just experimenting bovw approach with some random dataset. How should I label the dataset for the classification task?

Sky31 gravatar imageSky31 ( 2019-01-21 02:34:46 -0500 )edit

lets say the car, bike etc.

so, car:0, bike:1, tree:3, etc. (labels are just integer numbers)

and what are your "features" here ?

berak gravatar imageberak ( 2019-01-21 02:41:21 -0500 )edit

Features are the key-points which I extracted using SIFT/SURF.

I just need to give labels to each individual dataset lets say car dataset: 1, bike dataset: 2tree dataset: 3 and so on. ?

Sky31 gravatar imageSky31 ( 2019-01-21 03:30:42 -0500 )edit

AFTER creating a Bow Dictionary, and AFTER making Bow-features from your SIFT descriptors, you'll probably have 1 feature per image, and you need to label each feature for the classification

berak gravatar imageberak ( 2019-01-21 03:37:24 -0500 )edit

Correct me If I'm wrong.

Steps to execute:

Let's say I Input car dataset (1500 images) and create BOW dictionary, extract feature, therefore, It will output 1 feature per image. therefore, 1500 features of the car dataset.

similarly, bike dataset(1500 images) output 1500 features of the bike dataset.

Now I have to pick those features which I extracted from the dataset(s) and give a label to individual feature one by one and so on.


Sky31 gravatar imageSky31 ( 2019-01-21 04:09:01 -0500 )edit

all right !

the only thing i would change is: don't create a BowDictionary per dataset, but a large one for all your data (and classes)

berak gravatar imageberak ( 2019-01-21 04:14:37 -0500 )edit