099 Blob Track TOP

There’s been a FB thread about the Blob Track TOP in 099 and the removal of features in comparison to 088:
facebook.com/groups/touchde … 1169502552

I can speak for myself in saying that I do miss some features from 088’s Blob Track TOP. The background detection option in 088 in particular was very useful. In many cases, I’ve also had to do a lot more pre-processing of input textures to get usable results.

Is there any chance of revisiting this TOP?

Attached two versions
088 Vs 099.
Blob track in 099 seem to be bad bad node.
blob_cook_time099.toe (4.74 KB)
blob_cook_time088.toe (4.93 KB)

I’ll take a another pass at it, but one of the reason it got trimmed down was because we upgraded to OpenCV3.2, which had removed a lot of those features. Perhaps they are still available in an an external library.

Thanks for responding, Malcolm. Appreciate you taking another look at it!

Thank you. Any and all improvements to machine vision/people tracking features would be greatly appreciated.

I’ve discovered that the 099 top is actually more accurate but much slower. I tested this by feeding a TouchOUT top with source video(people tracking sample footage) from 088 to 099 with the same network and compared results.
099 TOP seems to perform fine at lower resolutions but scales up poorly, as in barakooda’s example. At lower resolutions they are comparable performance and one could argue that higher resolutions aren’t always necessary for blobs.
It behaves much better with blob interactions/collisions and seems to recognise blobs sooner. In 088 an ambiguous blob interaction seems to just madly swap blob positions etc momentarily sometimes.
Also after taking another look at the BG detection feature in 088 it doesn’t actually seem particularly useful in practice, it just doesn’t do a very good job, especially when “train bg” is re-enabled like in the docs “If you have the GPU/CPU power though you can leave it on even when the TOP is tracking blobs.” It just seems to ruin the background you have previously learned. I think the attraction of having an algorithm learning the background is to account for changing lighting conditions of the tracked area. This wont work in this case.
Something that both of them suffer from is not recognising blobs that touch the edge of the frame. These can be quite big/obvious but are not seen until they are entirely isolated in the frame.They both also suffer from bounds not bounding entire blob.
I hope there can be some improvements here without throwing the baby out with the bathwater.