search

beelabhmc / ant_tracker

Track ant movement in a lab setting using ML
http://hmcbee.blogspot.com/
1 stars 1 forks source link

run pipeline on test video #14

Closed aryarm closed 5 years ago

aryarm commented 6 years ago

see if you can just run the tracking part (no need to split)

potential issues: clear background

play around with matlab tracking settings to see whether we can improve tracking ex: reduce number of times ants can disappear but still be tracked ex: edit blob size threshold so that we can distinguish ants in the foreground (from the background)

mdonaldsonmatasci commented 6 years ago

Not bounding box but blob size threshold. Is the tracked object large enough to be an ant in the foreground?

aryarm commented 6 years ago

ah ok I've fixed the issue above

aryarm commented 6 years ago

test 1 results

ants seem to be tracked pretty well, even when they appear as small as in the test video. in fact, matlab seems to be capable of tracking ants much smaller than this. it often mistakenly tracked the shadows of ants against the sides of the box (which are much smaller than the ants themselves) unfortunately, ants underneath the bridge were also tracked

test 1 next steps and potential fixes

cropping out the bridge will prevent matlab from tracking the ant shadows as for ants that travel underneath the bridge, perhaps we could teach matlab what the shape of the bridge is? that way, it could identify when an ant is in one of the "holes" in the bridge and thus, make a decision about whether that ant is on the bridge or underneath it? this might even be possible with python alone using opencv polyfill to outline the bridge. for example, if we have the locations of ants in every frame, we can ask how many times the ants appear in a hole and calculate some sort of ratio for hole / not-hole appearance. the ants that are actually on the bridge will have a much lower hole / not-hold ratio

aryarm commented 6 years ago

new plan: test a single segment of the bridge and manually verify that counts are correct

aryarm commented 5 years ago

other problems identified

aryarm commented 5 years ago

The work for this issue was continued in #18, so I'm closing it for now.