We employ the same methodology for the evaluation as explained in the 2D case. Each point in the reference induces an entry in the confusion matrix, however, a buffered reference like in the 2D case will not be used. Special attention needs to be paid for non-delivered points. If missing points are actually part of a delivered class, they contribute to the False Negative count. If a class is completely ommitted it is indicated by -- in the confusion table (column), but if the respective points for the omitted class in the reference are labeled as an other class by the participants, the correctness of the latter one will decrease (row entries in the confusion matrix).

Vaihingen: 3D Labelling challenge



All quality measures except for overall are F1 scores in [%]. Note that in this table only a selection of classes - similar to the 2D case - are shown. The respective details-page shows the full statistics. Mouse over the column headings will display more information.

Abbrev. imp surf roof low_veg tree car Overall Strategy Details
IIS_185.082.865.848.264.069.9sD P
IIS_286.983.867.553.254.572.5sD P
IIS_385.386.073.057.140.271.4sD P
IIS_483.364.365.67.72.958.4sD P
IIS_583.083.459.756.133.568.0sD P
IIS_681.688.954.960.550.570.7sD P
IIS_785.090.965.275.657.976.2sD P
UM89.192.079.077.947.780.8sD P
HM_191.591.673.880.258.280.5sD P
WhuY62.781.959.268.741.564.4sD P
WhuY288.993.180.077.340.881.0sD P
WhuY390.193.481.478.063.482.3sD P
K_NN52.256.823.857.712.045.1sD P
K_LDA61.060.720.164.230.150.2sD P
K_QDA67.444.84.345.510.438.1sD P
K_RF64.151.78.253.316.541.5sD P
LUH91.194.277.583.173.181.6sD P
BIJ_W90.592.278.578.456.481.5sD P
NANJ90.993.677.777.151.780.7sD P
RIT_191.594.077.982.573.481.6sD P
NANJ291.293.688.882.666.785.2sD P
KIT80.081.857.970.944.168.1sD P
WUD1--87.4------84.0sD P
WhuY491.494.382.782.874.784.9sD
TUVI190.293.180.574.545.780.6sD