Lane Approximations

Because curves can be easier to handle than a few thousand pixels

Mean absolute distance

The original dataset metric. To get a feeling for the overall accuracy of the detector for each annotated lane segment.
Name All l1 l0 r0 r1 Comment
VSA SP 17.88 32.62 8.74 9.95 25.93 Trained on train and valid sets
VSA SP 18.47 32.84 8.57 11.56 26.51 Trained on training set only
VSA 19.00 34.77 8.74 10.79 27.84
Simple Mean Baseline 31.00 33.78 26.34 30.24 34.75 Within github repo

CULane Metrics

(Added early November) Accuracy metrics for detected lanes based on 30 pixel accuracy and an IoU greater or equal to 0.5
Name TP FP FN Precision Recall F1 Comment
LaneATT (ResNet-18) 68012 2161 7357 0.9692 0.9024 0.9346 Code and models are available at https://github.com/lucastabelini/LaneATT.
LaneATT (ResNet-34) 68495 2273 6874 0.9679 0.9088 0.9374 Code and models are available at https://github.com/lucastabelini/LaneATT.
LaneATT (ResNet-122) 68190 2239 7179 0.9682 0.9047 0.9354 Code and models are available at https://github.com/lucastabelini/LaneATT.
PolyLaneNet 66272 8302 9097 0.8887 0.8793 0.8840 Code and models are available at https://github.com/lucastabelini/PolyLaneNet.
Mean Baseline 917 82799 74452 0.0110 0.0122 0.0115 Not useful as baseline