Patricio Vela: Course Wiki

User Tools

Site Tools

Trace: • timing_interrupts • bluetooth_module
ece4580:module_classification

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revisionPrevious revision
Next revision		Previous revision
ece4580:module_classification [2017/01/25 20:19] – pvelaece4580:module_classification [2024/08/20 21:38] (current) – external edit 127.0.0.1
Line 1: Line 1:
-====== Classification ======+====== Image-Based Classification ======
  
 Classification involves obtaining a sample and determining to which class the sample belongs (from a finite set of classes).  The simplest form of classification is binary classification, where you must determine whether it is of class A or class B. The task ultimately boils down to determining identifying features of each class that are distinct from each other, then using those features to define a selection or classification policy.  Here, we will explore different ways to generate or discover the identifying features, also known as the //feature descriptor// or //feature vector// of the sample. Classification involves obtaining a sample and determining to which class the sample belongs (from a finite set of classes).  The simplest form of classification is binary classification, where you must determine whether it is of class A or class B. The task ultimately boils down to determining identifying features of each class that are distinct from each other, then using those features to define a selection or classification policy.  Here, we will explore different ways to generate or discover the identifying features, also known as the //feature descriptor// or //feature vector// of the sample.
Line 15: Line 15:
 This set of activities comes courtesy of Prof. Andrew Ng at Stanford university. We will be using his [[http://ufldl.stanford.edu/wiki/index.php/UFLDL_Tutorial|tutorials ]], which you will read and implement as you go through this learning module. This set of activities comes courtesy of Prof. Andrew Ng at Stanford university. We will be using his [[http://ufldl.stanford.edu/wiki/index.php/UFLDL_Tutorial|tutorials ]], which you will read and implement as you go through this learning module.
  
-**Week #1: Sparse Autoencoder** \\+__**Week #1: Sparse Autoencoder**__ \\
 Implement the 'sparse autoencoder' section. Implement the 'sparse autoencoder' section.
  
-**Week #2: Data Pre-Processing and Classification** \\+__**Week #2: Data Pre-Processing and Classification**__ \\
 Complete the `Vectorized implementation', 'Preprocessing: PCA and Whitening' and 'Softmax Regression' sections. With the complete of this step you should now have a digits classifier that does decently.  Here the feature space is the data itself, modulo the data normalization that might have been performed as the first step. Complete the `Vectorized implementation', 'Preprocessing: PCA and Whitening' and 'Softmax Regression' sections. With the complete of this step you should now have a digits classifier that does decently.  Here the feature space is the data itself, modulo the data normalization that might have been performed as the first step.
  
-** Week #3: Feature Learning and Classification.** \\+__** Week #3: Feature Learning and Classification.**__ \\
 Complete the 'Self-Taught Learning and Unsupervised Feature Learning' section for digit classification. Complete the 'Self-Taught Learning and Unsupervised Feature Learning' section for digit classification.
  
-** Week #4: Deep Neural Networks.** \\+__** Week #4: Deep Neural Networks.**__ \\
 Now, we are going to use the autoencoder repeteadly to create a deep network.  The learnt feature space from the first week will be analyzed to learn a feature space for that one (and maybe repeat the process). The idea being that each level will capture any special higher order relationships in the data.  At this point, your train classifier should have a classification accuracy of 98%. In prinicple, the deep neural network should be able to beat that.  Based on your results to date, please complete the 'Building Deep Networks for Classification' section. Now, we are going to use the autoencoder repeteadly to create a deep network.  The learnt feature space from the first week will be analyzed to learn a feature space for that one (and maybe repeat the process). The idea being that each level will capture any special higher order relationships in the data.  At this point, your train classifier should have a classification accuracy of 98%. In prinicple, the deep neural network should be able to beat that.  Based on your results to date, please complete the 'Building Deep Networks for Classification' section.
  
-** Week #5: Character Classification with Deep Neural Networks.** \\+__** Week #5: Character Classification with Deep Neural Networks.**__ \\
 Now you have a very nice classification system on digits. Can we classify on other similar data, such as characters, using the same approach?  The answer should be yes, but the bigger question is with what accuracy? Now you have a very nice classification system on digits. Can we classify on other similar data, such as characters, using the same approach?  The answer should be yes, but the bigger question is with what accuracy?
 Take your feature learning and classification training deep network procedure and train it with characters instead of digits. For characters, we would like to use [[https://github.com/brendenlake/omniglot/tree/master/matlab|Omniglot dataset]].  Let's do the following: Take your feature learning and classification training deep network procedure and train it with characters instead of digits. For characters, we would like to use [[https://github.com/brendenlake/omniglot/tree/master/matlab|Omniglot dataset]].  Let's do the following:
Line 48: Line 48:
 */ */
  
 +__**Week #1: Clustering to Define Words**__ \\
 +  - Study [[https://sites.google.com/site/dataclusteringalgorithms/k-means-clustering-algorithm|k-means clustering algorithm]] and the algorithmic steps for k-means clustering.
 +  - Download (or clone) the clustering skeleton code [[https://github.gatech.edu/fchu9/ECE4580_LearningModules/tree/master/objectRecognition/clustering|here]]
 +  - Implement k-means clustering algorithm working in RGB space by following the algorithmic steps. You are welcome to implement from scratch without skeleton code.
 +  - Test your algorithm on segmenting the image //segmentation.jpg// using k=3
 +  - Try different random initialization and show corresponding results.
 +  - Comment on your different segmentation results. 
 +
 +//Matlab Notes:// Matlab has several functions that can assist with the calculations so that you do not have to process the data in a for loops.  For example, there is **pdist2**, the mean function can process data row-wise or column-wise if specified properly, and there are ways to perform sub-assignments using either a binary array or the list of indices themselves.  You should be taking advantage of these opportunities to have compact code.
 +
 +__**Week #2: Object Recognition**__\\
 +
 +  - Study the [[https://www.youtube.com/watch?v=unCLMMsZAh8|bag-of-words]] approach for classification/Recognition task
 +  - We begin with implementing a simple but powerful recognition system to classify //faces// and //cars//.
 +  - Check [[https://github.gatech.edu/fchu9/ECE4580_LearningModules/tree/master/objectRecognition/simpleBoW|here]] for skeleton code. First, follow the README to setup the dataset and vlfeat library.
 +  - In our implementation, you will find [[http://www.vlfeat.org/index.html|vlfeat library]] very useful. One may use vl_sift, vl_kmeans and vl_kdtreebuild.
 +  - Now, use first 40 images in both categories for training.
 +  - Extract SIFT features from each image
 +  - Derive k codewords with k-means clustering in module 1.
 +  - Compute histogram of codewords using [[https://en.wikipedia.org/wiki/K-d_tree|kd-tree algorithm]] using vlfeat.
 +  - Use the rest of 50 images in both categories to test your implementation.
 +  - Report the accuracy and computation time with different k 
 +
 +
 +
 +__**Week #3: Spatial Pyramid Matching (SPM)**__\\
 +Usually objects have different properties across the spatial scales, even though they may appear common at one given scale.  Consequently, differentiation of objects is often improved by concatenating feature vectors or agglomerating feature descriptors that exist at different scales. This set of activities will explore how well that can work.
 +  - Study [[https://hal.archives-ouvertes.fr/file/index/docid/548585/filename/cvpr06_lana.pdf|Spatial Pyramid Matching]] which can improve BoW apporach by concatenating histogram vectors.
 +  - We will implement a simplified version of SPM based on your molude 2
 +  - First, for each traning image, divide it equally into a 2 × 2 spatial bin.
 +  - Second, for each of the 4 bins, extract the SIFT features and compute the histograms of codewords as in module 2
 +  - Third, concatenate the 4 histogram vectors in a fixed order. (hint: the a vector has 4k dimension.)
 +  - Forth, concatenate the vector you have in module 2 with this vector (both weighted by 0.5 before concatenated).
 +  - Finally, use this 5k representation and re-run the training and testing again.
 +  - Compare the results from module 3 and module 2. Explain what you observe.
  
 ----------------- -----------------
ece4580/module_classification.1485393584.txt.gz · Last modified: 2024/08/20 21:38 (external edit)