Experimental results of "Driving experience sharing method for end-to-end control of self-driving cars"

1. Training data examples from TORCS

* We collected the front-facing camera images from TORCS while a racing robot drove on 10 tracks. The driving information was collected at 10Hz. 

(a) normal driving

(b) the experience of returning to the lane center when a car is out of its lane.

2. Driving performance evaluation results

* We collected the front-facing camera images from GTAV while each CNN drove on the four courses. The driving information (steering command, yawrate, two dimensional positions of the car) was collected at 10Hz. However, the images were collected at about 7Hz because our PC system cannot handle 'saving images in real-time' over 7Hz.

(a) Ground-truth LDL versus LDL estimated by the proposed network

* Note that the red marks at the bottom of the video are artificially inserted.

(b) Driving performance comparisons: BL versus Prop.1

(c) Driving performance comparisons on course 1-4: SN v.s. DTN v.s. Prop.1 v.s. Prop.2