Point Cloud Segmentation Labels

Name every 3D point: chair, floor, wall. Segmentation paints the cloud part by part so robots grasp and rooms parse.

One cloud splits into colored parts with boundaries scored by overlap

Why Does This Exist?

Boxes tell a robot where a chair sits; grasping needs which points are legs versus seat. Segmentation exists for that per-point answer across semantic, instance, and part flavors. It drives indoor parsing, road labeling, and CAD part maps.

Networks that do it are KPConv, Point Transformer, and PointNet. Data basics are point clouds.

Think of It Like This

Coloring a statue by material

A restorer paints marble skin one color, bronze armor another, and cracks a third. Every chip gets exactly one label, and borders decide quality.

Points are chips and mIoU scores the paint job. The analogy stops at instances: statues have one armor while rooms hold ten identical chairs to separate.

How It Actually Works

Semantic heads output per-point class logits trained with cross-entropy over $13$ to $20$ classes. Instance heads add center votes plus clustering to split same-class objects. Part heads label subregions on single shapes. Score is mean IoU: for class $c$ , $\mathrm{IoU}_c = TP_c / (TP_c + FP_c + FN_c)$ , averaged over classes. UNet backbones with grid subsampling carry context from rooms to points.

Worked numbers

A chair class with $800$ true points predicted as $700$ chair plus $100$ table, and $100$ missed chair points, gives $\mathrm{IoU} = 700 / (700 + 100 + 100) = 0.778$ . Averaging $20$ ScanNet classes near $0.68$ is the reference operating point for strong locals.

Code

1def iou(tp: int, fp: int, fn: int) -> float:2    return tp / (tp + fp + fn)  # overlap over union3
4
5print(round(iou(700, 100, 100), 3))  # -> 0.778

Watch Out For

Trusting accuracy on imbalanced rooms

Floors dominate points, so $95$ percent accuracy can hide zero chair recall. The symptom is great logs with useless robots. Report mIoU and per-class scores, and sample rare classes harder during training.

The Quick Version

Segmentation names every point by class, instance, or part.
Semantic splits categories; instance splits objects; parts split shapes.
Mean IoU over classes is the score that matters.
Reference ScanNet mIoU sits near $0.68$ for strong local nets.
Chair demo with $700$ of $800$ right gives IoU $0.778$ .
Accuracy alone hides rare-class failures.