Different distance threshold for different classes

jkanjani · April 29, 2021, 1:41am

Is there a reason why annotations are thresholded at 50m before training? Also, why do we have a different distance threshold for different classes in the evaluation toolkit?

github.com

nutonomy/nuscenes-devkit/blob/master/python-sdk/nuscenes/eval/detection/configs/detection_cvpr_2019.json

{
  "class_range": {
    "car": 50,
    "truck": 50,
    "bus": 50,
    "trailer": 50,
    "construction_vehicle": 50,
    "pedestrian": 40,
    "motorcycle": 40,
    "bicycle": 40,
    "traffic_cone": 30,
    "barrier": 30
  },
  "dist_fcn": "center_distance",
  "dist_ths": [0.5, 1.0, 2.0, 4.0],
  "dist_th_tp": 2.0,
  "min_recall": 0.1,
  "min_precision": 0.1,
  "max_boxes_per_sample": 500,
  "mean_ap_weight": 5

This file has been truncated. show original

holger-motional · April 29, 2021, 3:40am

Hi, for training you can do what you want, these settings apply to evaluation. We manually reviewed samples for each class and decided until which distance they can still be reliably inferred. This is specific to our lidar’s range and density. For small objects like cones we set a much lower threshold than for large vehicle classes.

jkanjani · May 6, 2021, 6:39am

Thanks for the clarification.