search

ProjectSidewalk / sidewalk-cv-assets19

Repo for our ASSETS'19 paper applying ResNet to Project Sidewalk data
5 stars 6 forks source link

"Masking" of Sliding Window Area #5

Open galenweld opened 5 years ago

galenweld commented 5 years ago

One possible way to significantly improve our performance by reducing false positives is to adopt a "masking" approach, excluding regions of the panoramas from our sliding window computation that was can immediately rule out as not including sidewalk terrain. This approach was used in the Tohme paper:

screen shot 2019-01-29 at 15 26 05

Due to low-resolution and noise, we currently use 3D-point cloud data as a ground plane mask rather than as a feature to our CV algorithms.

galenweld commented 5 years ago

Currently, our system uses a "naive" mask, simply excluding anything more than X pixels above the horizon, and more than Y pixels below the horizon. Looking at our most common errors currently, they are most frequently placed in trees and foliage or houses adjacent to the street area, so some more sophisticated means of masking would be necessary to catch these rather than just using the depth data as a ground plane mask as in Tohme.

jonfroehlich commented 5 years ago

@galenweld, can you update this Issue with its current status or close it out... thanks!

galenweld commented 5 years ago

We do much better now than we did in January in terms of avoiding false positives, so this doesn't seem super urgent. That being said, it is definitely something that would be fun to experiment with more at some point.