Point Cloud Annotation Tools Compared: Open-Source and Paid LiDAR Tools

Compares Potato, CVAT, Supervisely, Segments.ai, Scale, Xtreme1 and SUSTechPOINTS for 3D LiDAR annotation, covering formats, cuboids, tracking, sensor fusion and agreement.

Most 3D annotation tools draw cuboids and track them across frames, and they differ mainly in which point cloud formats they read and whether sensor fusion costs extra. The sharper split is measurement. Across the other tools compared here, checked on 2026-09-24, every agreement statistic their documentation describes is scoped to 2D types, so a project needing chance-corrected agreement over 3D boxes has few options.

Point cloud annotation means placing labels on sparse, unevenly sampled 3D data, usually from LiDAR. How to annotate 3D point clouds covers the annotation problem itself, including rotation representation and why an axis-aligned agreement approximation misleads. This page compares the tools.

This page covers 3D only. Annotation Tools Compared covers every data type on one page.

Capabilities side by side

Every entry was read from the tool's own documentation on 2026-09-24. A dash means the documentation did not describe the capability, which is not the same as the tool refusing it.

Tool	Formats read	Point segmentation	Sequence tracking	Sensor fusion	License
Potato	PCD, PLY, KITTI `.bin`, LAS, XYZ, PTS	Yes, `segment_3d`	Yes	Camera calibration	Open source
Segments.ai	PCD, PLY, LAS 1.4 uncompressed, KITTI, nuScenes	Yes	Yes, with interpolation	Yes, 3D to 2D projection	Paid, Fusion tier
Supervisely	PCD, PLY, LAS, LAZ	Yes	Yes, episodes	Photo context projection	Enterprise
Scale	JSON or protobuf via API	Yes, per-point classes	Yes, frame series	Yes, reprojection	Commercial
CVAT	Velodyne `.bin`, PCD	-	Yes, interpolated tracks	Contextual images shown	Open source
Xtreme1	PCD	Claimed, workflow not documented	Yes, frame series	Yes, 2D and 3D fusion	Apache 2.0
SUSTechPOINTS	`.bin`, PCD	-	-	Multiple camera images	GPL-3.0
labelCloud	PCD, PLY, PTS, XYZ, BIN	Box-based only	-	-	GPL-3.0
point_labeler	KITTI `.bin`	Yes, point-wise	Multi-scan tiles	Reference images	MIT
Label Studio	No point-cloud object tag in the open-source tag reference	-	-	-	Enterprise interface
Labelbox	No point-cloud editor in the docs	-	-	-	-
V7	No point-cloud format listed	-	-	-	-

LAZ is the format that most often forces a conversion step. Supervisely reads it directly. Segments.ai's documentation lists LAS 1.4 uncompressed only, and Potato asks you to expand a LAZ file first with laszip or to install laspy[laszip].

Agreement over 3D labels

Three tools here report an agreement statistic at all, and all three scope it to 2D. CVAT's documentation is explicit, stating that "quality estimation only supports 2d tasks" and that it "supports all the annotation types except 2d cuboids". Supervisely's labeling consensus works with bounding boxes, masks, polygons, and image or object tags, and its score is a 0 to 100 percent pairwise similarity rather than a chance-corrected coefficient. Labelbox's consensus support table lists image, video, text, chat, audio, geospatial, document and HTML types, with no 3D entry.

The remaining tools' documentation does not mention annotator agreement over 3D labels. Treat that as absence of documentation rather than as a statement about the code.

The gap matters because agreement over geometry is not the same problem as agreement over labels. Two annotators who both mark a car disagree by degrees rather than categorically, so the measure has to match boxes to each other first and then ask whether the remaining variation exceeds chance. Potato computes that over 3D boxes using exact rotated 3D IoU. The point cloud guide warns against the axis-aligned approximation for the same reason, because an approximation reports disagreement that is an artifact of the measure.

Sensor fusion and what it costs

Projecting labels between the point cloud and the camera images is how an annotator checks work that looks correct in a bird's-eye view and is meters off along the view axis. The capability is common and the price attached to it varies widely.

Segments.ai projects 3D cuboids onto camera frames automatically and preserves the track ID across sensors, with sensor fusion gated to its Fusion tier. Supervisely projects photo-context colors onto the 3D points across every frame of an episode, with 3D LiDAR and sensor fusion listed under Enterprise and several of the assist tools behind a Cloud Max add-on. Scale reprojects cuboids into 2D camera views and offers a Lidar Linking task type for correcting those projections. CVAT displays contextual images alongside the cloud, and its documentation does not describe projecting labels between the two.

Potato takes camera calibration per item so that every 3D box projects into the images and can be checked in 2D, which is the same verification step without a tier attached.

Open-source options

Several of the tools here are open source and free to self-host, and they suit different jobs. CVAT is the most complete general platform, with cuboids and interpolated tracks. Xtreme1 is Apache 2.0 with 2D and 3D fusion and a frame-series tool, reading .pcd only. SUSTechPOINTS is GPL-3.0 and focuses on 9-degree-of-freedom box editing with automatic camera switching. point_labeler is MIT and is the tool SemanticKITTI was labeled with, doing point-wise semantic and instance labeling over KITTI sequences rather than boxes.

Label Studio's position is worth stating carefully. Its open-source tag reference lists no point-cloud object tag, and LiDAR is offered as an Enterprise interface described in HumanSignal's blog and press material rather than in reference documentation.

A 3D task in Potato

Potato's 3D tools are cuboid_3d, point_3d, polyline_3d and segment_3d, the last of which paints labels onto points directly.

yaml

annotation_schemes:
  - annotation_type: spatial_annotation
    name: traffic_objects
    description: "Draw a 3D box around each vehicle and pedestrian"
    tools:
      - cuboid_3d
    labels:
      - car
      - pedestrian
      - cyclist

tools selects which drawing modes the annotator gets, so a task that only needs boxes does not offer point painting. Level-of-detail rendering is on by default, which is what keeps a multi-million-point cloud interactive in a browser.

Frequently asked questions

The questions below match queries the site already receives for point cloud annotation.

Which point cloud annotation tool should I use?

Choose on formats and measurement rather than on drawing tools, because every tool here draws cuboids. If your data is LAZ, Supervisely reads it without conversion. If you need sensor fusion at no extra tier, CVAT shows contextual images and Potato projects boxes into calibrated camera views. If you have to report agreement between annotators on 3D boxes, check that the tool measures geometry rather than labels, since the documented agreement features in this comparison are scoped to 2D.

Is there an open-source 3D point cloud annotation tool?

Yes, and there are several with different strengths. CVAT, Xtreme1 (Apache 2.0), SUSTechPOINTS (GPL-3.0), labelCloud (GPL-3.0), point_labeler (MIT) and Potato are all free to self-host. point_labeler is point-wise rather than box-based and was used to label SemanticKITTI, while the others center on cuboids.