![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
What I've Been Doing
Why, I've been working on SPACE ROBOT code, of course. Only a few days to go!
Specifically, I've been working on the first movement estimation algorithm. It first simplifies an image by converting all the colors into 6-8 clusters. Then it combines all touching pixels of the same cluster color into regions. Choosing the largest regions between adjacent frames it can estimate motion by seeing how those regions have moved.
As you might expect, this is a real pain to debug, so I've developed a series of image output illustrating each step. Not only were they very useful in finding some subtle bugs and generally improving the algorithm, but they look pretty cool too.
So here is the output of a sample run. Top row is frame 1, second row is frame 2. The leftmost image is the original frame. (These were part of a 360 degree pan that I captured for testing purposes. That's![[livejournal.com profile]](https://www.dreamwidth.org/img/external/lj-userinfo.gif)
xmurf in his hat in our lab space.) The next image is the frame converted to only the 6 cluster colors. The third image is the frame with only the 'important' regions showing, the center of each marked with a white pixel. The final, combined image on the right shows the regions which have been matched together as the same, with lines connecting them. The motion estimation for these two frames was over 7 and down 4, which looks pretty good.
Specifically, I've been working on the first movement estimation algorithm. It first simplifies an image by converting all the colors into 6-8 clusters. Then it combines all touching pixels of the same cluster color into regions. Choosing the largest regions between adjacent frames it can estimate motion by seeing how those regions have moved.
As you might expect, this is a real pain to debug, so I've developed a series of image output illustrating each step. Not only were they very useful in finding some subtle bugs and generally improving the algorithm, but they look pretty cool too.
So here is the output of a sample run. Top row is frame 1, second row is frame 2. The leftmost image is the original frame. (These were part of a 360 degree pan that I captured for testing purposes. That's
xmurf in his hat in our lab space.) The next image is the frame converted to only the 6 cluster colors. The third image is the frame with only the 'important' regions showing, the center of each marked with a white pixel. The final, combined image on the right shows the regions which have been matched together as the same, with lines connecting them. The motion estimation for these two frames was over 7 and down 4, which looks pretty good.