|
Meeting #19
- Date: 24 - 2 - 2003 (Mon)
- Time: 4:15 - 5:20 pm
- Venue: Room 328, Chow Yei Ching Building
- Attendents: Dr. C. L. Yip, John Chan, John Fung, Tang Man Pan, Terry Tam
|
Minutes
- Fixes have been made in using circular blocks, applying step size for matching and wrong cooridnate system for output in our program.
- We suggested to do thresholding on pixel values before using block matching metrics (MAE & MSE). Dr.Yip said that this preprocessing could be done by first using histogram to find the distribution of pixel values, this would give us indications on the threshold value corresponding to a particular image.
- We suggested to do thresholding on block matching values (correlation, MAE and MSE) before finding motion vectors. Dr.Yip suggested us to consider also other measures. For example, the number of pixels being matched within a block.
- We had applied the Storke's Theorem such that the curl vectors in a region had been integrated. The size of the region was 180 pixels (step size equals 12 pixels and 15 motion vectors were considered) in diameter. Typical TCs in NWP has sizes of 3.7C lat. Assume the system is weaker in SCS and we only focus on core part but not outflow of the system, 180 pixels is enough in detecting its circulation. This is the reason why we chose this diameter. From our observation, the curl vectors at the center of the circle (potentially the eye) should be removed and we might try smaller step sizes to increase the accuracy.
- We would try to also apply another part of the Stokes's Theorem (integration of motion vectors along the boundary of the circle).
- We had used a perfect motion vector field to verify the correctness of the curl vector finding algorithm. Dr. Yip suggested us also tried the general cases. For example, consider motion vector fields with all vectors in the same direction or half of the vectors are in opposite directions when compared with the other half.
- In finding curl vectors, we had considered using three types of motion vectors (+,x,*), where the symbols described the locations of the vectors considered, in finding the curl vectors. Also, we had considered not only using the neighbours of a vector to find the curl values. Neighbourhood values (looping from 2-4) in finding curl vectors were used.
- Visualization (by 3D Grapher) of the relationship between lambda value, window size, dataset taken, start position and the error values in object tracking had been done. From our observation, the window size should be set to around 6, where the lamda value was about 0 to -0.1 for TCs with hourly reading (Maria was used in testing) and -0.5 to -0.6 for TCs with 6-hourly reading (Utor and York were used in testing). Dr.Yip suggested us to stop analysis on the variables dataset taken and the start position. And we should focus on the relationship between TC attributes, lambda values and window sizes, where more TCs of different types should be considered in testing.
- For postprocessing of motion vectors, minimum factor, maximum factor, mean filter, median filter and mid-point filter had been used, where dx and dy values varied from -5 to +7 after postprocessing.
- In order to make motion vectors more sensible, Dr.Yip suggested us to integrate the preprocessing (equalization and thersholding) and postprocessing (filtering) into the existing block matching algorithm as soon as possible.
- Next meeting with Dr. Yip will be on 3th March, 2003 (Mon).
Back to Index