Precise ego-localization is an important issue for Intelligent Vehicles. Geo-positioning with standard GPS often has localization error up to 10 meters, and is even sometimes unavailable due to "urban canyon" effect. It is therefore an interesting goal to design an affordable and robust alternative to GPS ego-localization. In this paper, we present 2 approaches for absolute ego-localization based on vision only, and not requiring previous driving on same street, by leveraging only pre-existing geo-referenced panoramas such as those from Google StreetView. Our first variant is based on Bag of visual Words + visual keypoints matching + bundle adjustment, and the other one uses direct pose regression computed by a deep Convolutional Neural Network (CNN) taking the query image as input. We have evaluated our 2 proposed variants using a real car. On around 1 km in a dense urban area, we obtained average localization errors of 2.8m with visual keypoints-matching + geometric computations, and of 7.7m with pose regression using pre-trained deep CNN. This shows that our proposed approaches are therefore potentially interesting complements or even alternatives to GPS localization.