Labs

Teaching: Radim Tyleček http://cmp.felk.cvut.cz/~tylecr1

Wednesday 16:15-17:45 in KN:E-128. Labs are given approx. every odd week (check course schedule).

Projects will be done in teams of 2 or 3 students. Team list is here. Each team can choose an application from the list provided below.

Task: Isolated word speech recognition for a small vocabulary

Data: audio data containing isolated words

Model: Each word of the vocabulary is modelled by a Markov model. The hidden states can be thought to represent phonemes. The features are spectral vectors obtained by local Fourier transform

Learning: Realize unsupervised learning i.e. only the word class is known for each sample in the training data.

Available data: audio data containing numbers between 0 and 30 spoken by different speakers.

Remarks: The learning data contain sequences of spoken numbers with short pauses in between. Think how to segment them into isolated words automatically.

ShATR Corpus

Task: Text line recognition in license plates

Data: Uniformly scaled images of Czech license plates

Model: Images are considered as sequences of characters and white spaces, where each character has a known (fixed) width, whereas white spaces have varying width. More specific: The image is considered as a sequence of columns with names. This sequence is to be modelled as a Markov chain. The appearance model can be assumed (in the simplest case) as follows: Given the name of the column, the grey values for each pixel are e.g. normally distributed, where the mean and variance for each pixel are (unknown) parameters.

Learning: Realize two variants of learning

a) Supervised learning. In this case the learning data are fully annotated, i.e. the sequence of characters and their positions are known.

b) Semi-supervised learning. Here only the character sequence but not their positions is known.

Available data: Fully annotated, uniformly scaled images of Czech license plates in Matlab format.

Task: Estimating the epidermis-dermis boundary in OCT images

Data: Cross-section images of skin captured by Optical Coherence Tomography (OCT)

Model: The background-epidermis and epidermis-dermis boundaries are considered as sequences of height values and should be described by a Markov chain. The image is considered as a sequence of columns. Find a model which relates the two height values with the image of a column.

Learning: Combine supervised and unsupervised learning for the estimation of the model parameters.

Available data: partially annotated OCT images of the skin of different patients.

Chosen application will be solved by modeling an appropriate (hidden) Markov model on a chain.

Teams are asked to present their plan at the beginning of Lab 4 in approx. 10 minutes. The plan should include answers to the following questions:

• Features
  1. What will my features represent?
  2. How can I generate feature vectors from the input data?
• Probability model
  1. How many hidden states will the model have?
  2. What the hidden states correspond to?
  3. How will a typical sequence look like? (Think about transition probabilities.)
  4. How can I calculate the emission probability for a given feature vector?
• Learning
  1. What parameters will have to be learnt in supervised case?
  2. What parameters will have to be learnt in unsupervised or semi-supervised case?
• Recognition
  1. How will be the resulting sequence transformed into final output?

Students are expected to code the core inference and learning algorithms from the scratch in language of their choice (C/C++/C#, Java, Matlab). On the other hand, any available Open Source software can be used for the infrastructure/environment part (i.e. data reading, pre-processing, output presentation).

Teams are expected to submit a written report (5-10 pages) describing their approach to solve the chosen task. The report will be evaluated focusing on the correctness of the solution.

Write directly about your approach, do not explain things given in the lectures.

Basic structure

1. Header: name, authors, date
2. Short introduction describing the problem
3. Model description - states, probability models
4. Method description - learning, recognition
5. Results - figures, error rates
6. Conclusion - achievements, problems

Submission

Send an email with PDF to Radim Tylecek. Deadline: 31.12.2010 23:59

Each group has to present its results (5-10 mins) at a presentation session organized at the end of the semester. A live demonstration of the resulting application should be a part of the presentation. The presentation evaluation will also include the real performance of the implemented algorithm.

See course schedule.

Student's team must obtain at least 50% in the written and oral presentation to obtain lab credit (zápočet).