Skip to content

aghagol/loop-detection

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Loop detection using deep net features

AlexNet feature vectors for loop closing in SLAM

It has been proposed that feature vectors from AlexNet are robust with respect to appearance changes:

N. Sunderhauf, F. Dayoub, S. Shirazi, B. Upcroft, and M. Milford, On the performance of convnet features for place recognition, arXiv preprint arXiv:1501.04158, 2015.

Look in the folder PR_bench for our implementation of the above paper over the Gardens Point dataset.

Demo:

We use feature vectors from the conv3 layer of AlexNet to match each frame of the testing video with a frame in the reference video that is captured on a different day. Videos were captured using a Google Nexus 5 phone. GPS is recorded simultanously using a dashcam Andriod app. I put a cheap ($5) fisheye lens on Nexus 5 to increase its field of view.

  • Matching video frames from a cloudy day to video frames of a sunny day and predicting the location. On the map, big blue circle corresponds to the GPS location that is recorded over each test video frame (ground-truth) and the green star corresponds to the GPS location of the reference video frame (prediction).

IMAGE ALT TEXT HERE

About

AlexNet feature vectors for loop closing

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages