Perception weights and config files

This zip contains all necessary files for running the vision_cone_detector package, including weights and config files that must be specified in vision.yaml.

• cones5.weights: trained for yolov4-tiny, 1920x1088 input images from FSOCO public dataset, with 5 custom anchors for:

  • yellow_cone
  • blue_cone
  • orange_cone
  • large_orange_cone
  • unknown_cone

• keypoint_weights.pth: trained for our modified ResNet to detect 7 keypoints of each cone, using the MIT-RekNet public dataset. We currently use only one vertex, but all are inferred.