We talk about several topics.
1. Calibration of the frames.
– Lowe uses the reproduction error to give an idea of how good calibrated that is. An optimisation coluld make it easy.
2. Probabilistic data association.
– The final update matrix is composed of the cog of not using any association in addition to the cov of using only kalman filter. Plus another term which represents the data association, called P tilde.
3.Kalman filter is not going to give me a more accurate estimate, there is a trade off between robustness and accuracy. It will smooth the sequence. So an homography that is wrong, would be less wrong.
4. Now it is time to introduce data association. So we have points in an image. The points reproject in the second one. Then, if we use (J)PDA to associate this points, then we would have an estimate of the right point locations. Then, we can use those to compute our new homographies.
5. I tried to get an implementation of multi band blending. Something not working properly. Tom suggested using ‘enblend’ to check how should the output be.
6 . Take a picture of a scene, and go closer and try to mosaic it. It has the problem of the calibration of the camera. But visually is ok.