Demo | VL3D++

Home

Raw point cloud—Original unlabelled point cloud

vsFinal result

Loading visualisation…

◀▶

Raw point cloud

Final result

⇔Drag to compare

Summary

80.2%

Overall Accuracy

70.6%

Precision

66.1%

Recall

66.3%

F1 Score

Full metrics

IoU?

Intersection over Union (macro)Overlap between prediction and ground truth per class, averaged unweighted. Reference values are lower in 3D point clouds than in 2D imagery.0–100% · >40% acceptable · >60% good in 3D

56.1%

wP?

Weighted PrecisionPer-class precision weighted by support (number of points). Reflects practical performance weighted by the real dataset distribution.0–100%

70.6%

wR?

Weighted RecallPer-class recall weighted by support. In point-by-point classification with mutually exclusive classes, its numerical value equals Overall Accuracy.0–100% · Expected value ≈ OA

66.1%

wF1?

Weighted F1 ScoreHarmonic mean of wPrecision and wRecall, weighted by class support. More informative than OA with imbalanced classes.0–100%

66.3%

wIoU?

Weighted IoUPer-class IoU weighted by support. Gives more weight to larger classes. Macro IoU is more sensitive to minority-class performance.0–100%

56.1%

MCC?

Matthews Correlation CoefficientCorrelation between predictions and ground truth. Considers all confusion-matrix elements. Reliable with imbalanced classes.-100 to 100 · near 0 = no correlation · 100 = perfect

69.9%

Kappa?

Cohen's KappaAgreement between model and ground truth corrected for chance agreement. More robust than OA on imbalanced datasets.0–100% · <40% poor · 60–80% good · >80% very good

69.1%

Macro vs Weighted

Metric

Macro

Weighted

Precision

70.6%

Recall

66.1%

66.3%

IoU

56.1%

Macro = unweighted arithmetic mean per class. More sensitive to minority-class performance. Weighted = weighted by class support (number of points). Better reflects practical impact on the real dataset.

Layer guide

Raw point cloud
3D point cloud without labels. Colour encodes LiDAR return intensity or normalised elevation.

Error mask
Per-point classification accuracy vs ground truth. Red = likely error. White/grey = correct classification.

Uncertainty
Entropy of the predicted class probability distribution per point. Yellow = maximum doubt. Purple = high confidence.

Ground truth
Expert-validated semantic labelling. Reference against which all model metrics are computed.

Final result
Model prediction over the entire cloud. Each point assigned its semantic class. Main VL3D++ output.

Raw point cloud—Original unlabelled point cloud

vsFinal result

Loading visualisation…

◀▶

Raw point cloud

Final result

⇔Drag to compare

Summary

80.2%

Overall Accuracy

70.6%

Precision

66.1%

Recall

66.3%

F1 Score

Full metrics

IoU?

56.1%

wP?

Weighted PrecisionPer-class precision weighted by support (number of points). Reflects practical performance weighted by the real dataset distribution.0–100%

70.6%

wR?

66.1%

wF1?

Weighted F1 ScoreHarmonic mean of wPrecision and wRecall, weighted by class support. More informative than OA with imbalanced classes.0–100%

66.3%

wIoU?

Weighted IoUPer-class IoU weighted by support. Gives more weight to larger classes. Macro IoU is more sensitive to minority-class performance.0–100%

56.1%

MCC?

69.9%

Kappa?

Cohen's KappaAgreement between model and ground truth corrected for chance agreement. More robust than OA on imbalanced datasets.0–100% · <40% poor · 60–80% good · >80% very good

69.1%

Macro vs Weighted

Metric

Macro

Weighted

Precision

70.6%

Recall

66.1%

66.3%

IoU

56.1%

Layer guide

Raw point cloud
3D point cloud without labels. Colour encodes LiDAR return intensity or normalised elevation.

Error mask
Per-point classification accuracy vs ground truth. Red = likely error. White/grey = correct classification.

Uncertainty
Entropy of the predicted class probability distribution per point. Yellow = maximum doubt. Purple = high confidence.

Ground truth
Expert-validated semantic labelling. Reference against which all model metrics are computed.

Final result
Model prediction over the entire cloud. Each point assigned its semantic class. Main VL3D++ output.