Place Recognition with ConvNet Landmarks: Viewpoint-Robust, Condition-Robust, Training-Free