Background removal improvement

berkeleyauv / perception

Autonomous Underwater Vehicles at Berkeley's perception stack

http://urobotics.berkeley.edu/

9 stars 1 forks source link

Open nLevin13 opened 3 years ago

nLevin13 commented 3 years ago

Ideas:

nLevin13 commented 3 years ago

Switch between the two algorithms instead of using both at once
Interface with behavior tree from controls to figure out what we’re doing at the moment
High frequency changes -> switch to another algorithm
Gpu implementations for opencv for more speed

KDharmarajanDev commented 3 years ago

[x] Cleaned up unused imports/code in combined algorithm file 83a2186e24566bdec6857e2115a3e44b8ac1cd0f
[x] Tested upscaling to 25% after running algorithms (Result is that upscaling to 25% doesn't affect FPS)

KDharmarajanDev commented 3 years ago

Worked on switching algorithm between two sub algorithms. Current problems

Large contours tend not to change their centroids rapidly, making it so that the algorithm doesn't switch to a better one.
Sometimes there is rapid change of contours, leading to a significant physical difference in object detection (and it also goes back and forth very rapidly)
Smaller image representation via pixel averaging for better performance?
Need to test timing with switching algorithms

nLevin13 commented 3 years ago

KDharmarajanDev commented 3 years ago

[x] Add multi-contour support
[x] Tune large area heuristic component
[x] Getting codebase ready for nondebug use
[ ] returning list of named contours
[x] don't do calculations that are only needed on debug mode (namely everything we need to make the debug frames)

KDharmarajanDev commented 3 years ago

[x] Integrate saliency detection/background subtraction into roulette wheel color detector
[ ] Add dice bounding boxes
[ ] Fix num_contour reset
[x] Interpolation between frames where we don't have accurate contours