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.
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).
