•1 min read•from Machine Learning
Showcase: geolocating a dashcam video without GPS, only from the footage [P]
Sharing a project I have been working on called Third Eye. It does visual geolocation. Given a video, it figures out where it was filmed using only the image content, and draws the route on a map.
Pipeline in short:
- per frame place recognition against a street imagery index
- a trajectory search that stitches the frames into one coherent path
- a geometric verification step to catch false matches
per frame confidence so weak frames are flagged, not faked
I ran it on real dashcam footage and it traced the route quite well. Cross domain matching like this is genuinely hard, so a fair amount of the work went into making it honest about uncertainty.
Keen to hear feedback on the matching and trajectory side.
Video Demo: https://youtu.be/U3sItFlvq6E?si=-KJrwb0gSlk-GxVH
The Index was covering a 12KM2 Area around NYC.
[link] [comments]
Want to read more?
Check out the full article on the original site
Tagged with
#rows.com
#real-time data collaboration
#real-time collaboration
#visual geolocation
#dashcam footage
#place recognition
#street imagery index
#trajectory search
#geometric verification
#cross-domain matching
#uncertainty
#per frame confidence
#machine learning
#image content
#route tracing
#map
#NYC
#index
#frames
#geolocation