Visual DNN Planner

This repo generates path proposal labels from LIDAR pointclouds, pose estimates and mono images. The labels can be used to train a CNN autoencoder to propose valid path for autonomous car applications. It is based on the the following work:

D. Barnes, W. Maddern and I. Posner: "Find Your Own Way: Weakly-Supervised Segmentation of Path Proposals for Urban Autonomy,"

Check out preliminary results here!

Instructions to run

Please modify camera_lidar_extrinsic.yaml to reflect the proper extrinsic calibration parameters from the camera to the LIDAR. The file udacity_sdc_center_cam.yaml should be changed to values for the intrinsic parameters of your camera.

To launch the label generator against the Udacity ROS bag datasets here, please run:

roslaunch visual_dnn_planner label_generator.launch
rosbag play <bag name>.bag

The output images are saved under /tmp by default.

TODOs

1. Include SegNet CNN training and testing scripts after cleanup