Transcriptions
Note: this content has been automatically generated.
00:00:04
nice place and i work on detecting unusual obstacles
00:00:09
in traffic images in order to improve the safety of
00:00:12
set of driving cars uh let's say our car 'cause it come or how does it know what it sees
00:00:19
common approach is to assign label every peak so the pet pixels belong to
00:00:24
the route between once the trees and the dark it was to the cox
00:00:29
this is called semantic segmentation it is often done made you run networks
00:00:35
they are very you they work very well but
00:00:38
they require many traded examples thousands of images don't labels
00:00:44
we had those examples for common classes such as people cars buildings rolls
00:00:51
and so on however when we encounter red objects such as an aimless
00:00:57
or last cargo which the network is never seen training
00:01:01
it fails to notice the objects or produces nonsense outputs
00:01:05
i want to take these cases and usually working the ten unknown objects press it's
00:01:11
my approach is to reconstruct the image from the labels
00:01:15
then i found find how it differs from the original image
00:01:20
if an object is incorrectly they but it will not look to send reconstruction
00:01:26
these differences show us what the labels are wrong
00:01:31
here is an example we some laws boxes i hope this work will improve the safety of sort
00:01:36
of driving cars as well as the tact interesting training cases for other are great it's thank you