I want to train a model on 3D point cloud data. What is the best LiDAR or 3D point cloud annotation tool?

If your goal is training a model on LiDAR/3D point-cloud data, my top pick for most ML teams would be Segments.ai. It is purpose-built around 3D point clouds rather than treating 3D as an add-on to an image-labeling system.

My shortlist

Current comparisons similarly put Kognic/Segments.ai toward the purpose-built end of the spectrum, while CVAT is a strong open-source/self-hosted option.

Why I'd choose Segments.ai

For actual 3D ML training, the important thing isn't simply "can it draw a 3D box?" You want the annotation tool to support the type of labels your model will consume.

Segments.ai supports:

• 3D cuboids
• Semantic segmentation
• Instance segmentation
• Panoptic segmentation
• Polygons, polylines and keypoints
• LiDAR + camera data
• Temporal/sequence annotation
• Object tracking and interpolation
• Large point clouds through 3D tiling
• Python SDK/API integration

Its segmentation interface lets annotators select points using brush, polygon, or box tools, which is particularly useful if you're training a point-wise semantic/instance segmentation model rather than just a 3D object detector.

The camera/LiDAR fusion is another major advantage. You can upload calibrated camera images and inspect the 3D annotations overlaid on the imagery.

When I'd choose CVAT instead

If you're comfortable running infrastructure yourself and want to avoid paying for a commercial annotation platform, I'd seriously consider CVAT.

The Community edition is free and can be deployed on your own infrastructure. It supports 3D .pcd and .bin data, annotation workflows, QA, and automated annotation.

The tradeoff is that Segments.ai generally gives you a more specialized 3D/LiDAR workflow out of the box, whereas CVAT is attractive when flexibility, self-hosting, and cost matter most.

What I'd recommend based on your model

• 3D object detection → Segments.ai or CVAT; annotate oriented 3D bounding boxes.
• Semantic segmentation → Segments.ai; its point-level segmentation workflow is particularly well suited.
• Instance segmentation → Segments.ai.
• Autonomous driving / robotics with LiDAR + cameras → Segments.ai or Kognic.
• Large temporal LiDAR sequences → Segments.ai/Kognic, because tracking and interpolation can save enormous annotation time.
• Research project / startup / limited budget → CVAT.
• Need a self-hosted solution for sensitive data → CVAT is an especially compelling starting point.
• Need annotation + dataset management + enterprise ML workflow → Encord is worth evaluating.

If you tell me what kind of point cloud you have (e.g. autonomous-driving LiDAR, indoor scans, aerial LiDAR, industrial 3D scans, RGB-D) and what model you're training (PointNet++, Point Transformer, 3D YOLO, BEVFusion, segmentation, etc.), I can narrow this down to the best 2–3 tools and the exact annotation format/schema I'd use.

When I'd choose CVAT instead

If you're comfortable running infrastructure yourself and want to avoid paying for a commercial annotation platform, I'd seriously consider CVAT.

The Community edition is free and can be deployed on your own infrastructure. It supports 3D .pcd and .bin data, annotation workflows, QA, and automated annotation.

Choosing the "best" 3D point cloud annotation tool depends heavily on your budget, whether you need 3D cuboids (object detection) or semantic segmentation (point-level labeling) , and if you require sensor fusion (syncing LiDAR data with 2D camera feeds).

Here is a breakdown of the best 3D point cloud annotation tools categorized by project type:

Best Open-Source & Self-Hosted Tools (Free)

• CVAT (Computer Vision Annotation Tool): Best open-source option for 3D cuboids . It supports native 3D point cloud formats (like PCD and LAS), allowing you to draw bounding boxes, adjust orientations, and trace objects. However, it lacks advanced point-level semantic segmentation for 3D.
• CloudCompare: Best for manual/semi-automated geospatial data . While traditionally an engineering tool, its "Cloud Layer" and algorithmic segmentation features (like RANSAC and CSF filters) make it incredibly powerful for manually classifying large terrain, building, or geological datasets before training a model.
• 3D-BAT (3D Bounding Box Annotation Tool) : A specialized, open-source web-based tool explicitly designed for labeling point clouds in autonomous driving setups.

Best Commercial Platforms for Rapid ML Pipelines

• Segments.ai: Best for 3D semantic segmentation and robotics . It features highly responsive point-level labeling and smart click-to-segment automation, making it a top choice for teams training complex perception models.
• BasicAI: Best for speed and AI-assisted pre-labeling . It offers advanced "auto-annotation" tools that leverage machine learning to automatically track objects across sequential point cloud frames, drastically lowering manual annotation hours.
• Supervisely: Best for end-to-end MLOps integrations . Supervisely allows you to plug your custom neural networks directly into the annotation toolbox to iteratively pre-label data and retrain your model within the same ecosystem.
• iMerit (Ango Hub) & Encord: Best for enterprise-level sensor fusion . If your training data pairs LiDAR with radar and multiple synchronized camera arrays, these platforms excel at projecting 2D camera boxes into 3D cuboids smoothly.

How to Choose the Right Tool

To help narrow this down, what specific format is your point cloud data in (e.g., .pcd, .las, .bin), and are you looking to do 3D bounding boxes or point-by-point segmentation?

Choosing the "best" LiDAR or 3D point cloud annotation tool depends heavily on your budget, whether you need AI-assisted/auto-labeling features, and if you are working with single frames or sequential LiDAR sweeps (like autonomous driving sequences).

The top-tier 3D point cloud annotation tools available are categorized below by use case and platform type:

1. Best Open-Source & Free Tools

• CloudCompare: A legendary open-source 3D point cloud editing and processing software. While not a dedicated AI-annotation platform with sleek team management, it is unrivaled for manual 3D data manipulation, segmentation, slicing, and scalar field comparisons. Perfect for research or heavy geometry processing.
• CVAT (Computer Vision Annotation Tool): Originally by Intel, CVAT is a robust, feature-complete open-source tool that supports 3D cuboid/bounding box annotation for LiDAR and radar point clouds. It handles multi-sensor data well and is great if you want to self-host a collaborative tool for free.

2. Best Commercial & Enterprise Platforms (AI-Assisted)

• Supervisely: Offers a powerhouse cloud-based 3D point cloud toolbox. It features powerful auto-labeling, object tracking across sequential frames, customizable ground segmentation, and the ability to integrate custom neural network pipelines for automated pre-labeling.
• Labelbox: An industry standard for enterprise data training. Labelbox supports complex 3D sensor fusion data (LiDAR + camera sync), custom labeling interfaces, and strong workflow/team orchestration tools for scaling up large datasets.
• V7 (V7 Darwin): Known for top-tier automation and AI-assisted data labeling. V7 handles 3D point clouds seamlessly and accelerates annotation through smart tooling and model-in-the-loop workflows.

3. Best for Research & Custom Python Pipelines

• Open3D: An open-source modern library that supports rapid development in 3D data processing. While it is a programming library rather than a standard GUI annotation tool, many research teams build custom, lightweight visualization and labeling scripts using Open3D's GUI features.

Which one should you pick?

• Choose CloudCompare if you need deep geometric analysis or free, high-precision manual editing.
• Choose CVAT if you need a free, collaborative, self-hosted web tool for a team.
• Choose Supervisely or Labelbox if you have budget, large volumes of data, and urgently need AI auto-labeling and sensor fusion (LiDAR + video) to cut down manual hours.

To help narrow down the right choice, tell me:

• Are you working with isolated 3D scans or sequential LiDAR video frames (like autonomous driving)?
• Do you need sensor fusion (matching camera images with LiDAR point clouds)?
• Is your preference for an open-source/self-hosted tool or a managed commercial platform?

If your goal is to train a model on LiDAR/3D point clouds, my shortlist would be:

My recommendation

If you're building your own 3D ML model, I'd start with Supervisely.

It has particularly good support for 3D point clouds from LiDAR/RADAR, point-level segmentation, 3D bounding boxes, sequences, camera context, and AI-assisted labeling. It also has pretrained 3D models such as CenterPoint, PV-RCNN, PointPillars, etc., which can be used to generate initial annotations rather than labeling everything manually.

If you're specifically working with autonomous-vehicle-style LiDAR sequences, I'd seriously consider Segments.ai. It is purpose-built around point-cloud/multisensor annotation and is one of the prominent dedicated options in current LiDAR tooling comparisons.

When I'd choose CVAT

If you want open source, self-hosting, and maximum control, CVAT is an excellent starting point. Its current platform supports 3D point clouds, including .pcd and .bin, 3D object annotation, cuboids, and tracking objects through point-cloud sequences.

The important distinction is that CVAT is more of a general-purpose annotation platform, whereas Supervisely and Segments.ai are more compelling if 3D/LiDAR is at the center of your workflow.

One important question

The "best" tool changes substantially depending on what labels you're training for:

• 3D object detection → 3D bounding boxes/cuboids
• Semantic segmentation → every point gets a class
• Instance segmentation → every point gets an object ID
• Autonomous driving → boxes + tracking + LiDAR/camera fusion
• Robotics → objects/parts + 6D pose
• Geospatial/LiDAR mapping → buildings, vegetation, ground, poles, terrain, etc.

If you tell me what your point clouds are (e.g. .las, .pcd, .bin, Ouster/Velodyne), what model you're training (PointNet++, PointTransformer, YOLO3D, BEVFusion, etc.), and whether you need bounding boxes or point-level labels, I can narrow this down to the best 2–3 tools and tell you exactly what I'd use.

If you're building your own 3D ML model, I'd start with Supervisely.

It has particularly good support for 3D point clouds from LiDAR/RADAR, point-level segmentation, 3D bounding boxes, sequences, camera context, and AI-assisted labeling. It also has pretrained 3D models such as CenterPoint, PV-RCNN, PointPillars, etc., which can be used to generate initial annotations rather than labeling everything manually.

If you're specifically working with autonomous-vehicle-style LiDAR sequences, I'd seriously consider Segments.ai. It is purpose-built around point-cloud/multisensor annotation and is one of the prominent dedicated options in current LiDAR tooling comparisons.

When I'd choose CVAT

If you want open source, self-hosting, and maximum control, CVAT is an excellent starting point. Its current platform supports 3D point clouds, including .pcd and .bin, 3D object annotation, cuboids, and tracking objects through point-cloud sequences.

If your goal is training a 3D ML model from LiDAR/point-cloud data, my top recommendation today is Supervisely.

My ranking

🥇 Best overall: Supervisely

Supervisely is particularly attractive because it isn't just a labeling UI. It provides a workflow from annotation → AI-assisted labeling → dataset management → model training.

For 3D data, it supports:

• LiDAR and RADAR point clouds
• 3D bounding boxes/cuboids
• Point-level segmentation
• 3D object tracking across sequences
• Ground segmentation
• Camera + LiDAR sensor fusion
• 2D-to-3D annotation transfer
• AI-assisted cuboid generation
• Pretrained 3D detection models
• PCD, PLY, LAS, LAZ and other formats
• KITTI 3D and ROS Bag export/import

It can also handle very large clouds—Supervisely says its browser-based interface supports up to ~50 million points per frame.

The really interesting part for model training is the automation. You can use pretrained models such as CenterPoint, PointPillars, PV-RCNN, 3DSSD and others to generate initial labels and then have humans correct them.

That creates a much better workflow than manually labeling every object from scratch.

---

🥈 Kognic — if you're doing autonomous vehicles

If your point clouds come from cars, robotics, ADAS, or multi-LiDAR sensor rigs, I'd seriously evaluate Kognic.

It's specifically designed around automotive sensor data and supports LiDAR + camera + radar, 3D cuboids, segmentation, polylines, multi-LiDAR setups, temporal interpolation, pre-labeling and automated QA.

Its big advantage is production-scale QA. Kognic advertises 90+ automated checker apps for detecting annotation errors and has an API-first workflow for integrating annotation into ML pipelines.

I'd choose Kognic over Supervisely if you're building something like:

10,000+ hours of autonomous-driving data → annotation workforce → automated QA → production perception model.

---

🥉 CVAT — best open-source option

CVAT is worth considering if you want something you can host yourself and don't need as sophisticated a 3D workflow.

It supports 3D annotation for things such as vehicles, buildings and landscapes, including 3D cuboids and tracking, and supports .pcd and .bin point-cloud data.

The tradeoff is that its 3D tooling is less specialized than Supervisely/Kognic. I'd pick CVAT when cost, open-source deployment, or customization matters more than having the best 3D annotation experience.

---

What I'd choose for your project

It depends heavily on the type of point cloud:

Object detection
→ Supervisely

Semantic/instance segmentation
→ Supervisely

LiDAR sequences / tracking
→ Supervisely or Kognic

LiDAR + camera fusion
→ Supervisely or Kognic

Autonomous driving at production scale
→ Kognic

Research project / university / self-hosted
→ CVAT or Supervisely

Large outdoor scans / mapping / geospatial LiDAR
→ Supervisely, though I'd also consider specialized GIS/LiDAR tooling depending on what you're labeling.

One important consideration

For training a model, I wouldn't choose an annotation tool based primarily on how nice the 3D viewer looks. I'd evaluate:

1. Can it generate the exact labels your model needs?
2. Can it propagate labels across frames?
3. Can you import your model's predictions as pre-labels?
4. Can you export directly into your training format?
5. Can you programmatically ingest/export data through an API?
6. Can multiple annotators work with QA/versioning?

On those dimensions, Supervisely is probably the strongest all-around choice for a new 3D ML project. Its current tooling specifically combines single-frame and sequential point clouds, AI labeling, tracking, 2D/3D synchronization, and dataset management.

If you tell me what kind of LiDAR data you have (e.g. autonomous vehicle, drone, Ouster/Velodyne, terrestrial scanner, indoor robotics) and what model you want to train (3D detection, segmentation, classification, etc.), I can narrow this down to the best 2–3 tools and recommend an annotation format/schema for your dataset.

The "best" LiDAR or 3D point cloud annotation tool depends heavily on your budget, whether you require open-source (self-hosted) or enterprise-grade (cloud/managed) software, and if you need sensor fusion (syncing LiDAR with RGB camera feeds).

The top options on the market for machine learning are organized below by category:

1. Best Open-Source & Self-Hosted Tools (Free)

• • Best For: Teams wanting a robust, highly customizable, and completely free self-hosted solution.
  • Key Features: Supports 3D bounding boxes (cuboids), trajectories, and point cloud sequences. It has active community support, user management, and can be integrated directly into local machine learning pipelines.
• CloudCompare
  • Best For: Advanced geospatial and research-focused 3D point cloud processing.
  • Key Features: An open-source desktop software powerhouse. While less of a collaborative "annotation pipeline" tool and more of a technical processing suite, it handles massive point clouds effortlessly, supports custom scalar fields, and offers advanced manual and semi-automated geometric segmentation (like RANSAC shape detection).

2. Best Enterprise & Commercial Platforms (AI-Assisted & Scalable)

• • Best For: High-performance spatial AI and complex multi-modal/sensor fusion pipelines.
  • Key Features: Widely regarded as a top-tier platform for spatial annotation. It treats data labeling as a complete system, supporting cuboids, segmentation, and tracking across temporal sequences, alongside advanced quality assurance metrics and deep LiDAR + camera fusion.
• • Best For: Advanced semantic/instance segmentation and 2D-to-3D sensor fusion workflows.
  • Key Features: Exceptional web-based tools that let you annotate point clouds directly via synced 2D camera images (using familiar brush masks and smart tools that auto-convert into 3D labels). It features strong auto-tracking and customizable neural network integrations for automated pre-labeling.
• • Best For: Processing massive, high-resolution multi-frame LiDAR datasets.
  • Key Features: Handles huge point volumes (over 150 million points across hundreds of frames) smoothly. Offers an intuitive UI with strong AI-assisted pre-labeling tools and options for hybrid managed workforce services.
• • Best For: Streamlined, fast robotics and autonomous driving workflows without unnecessary bloat.
  • Key Features: Specializes in fast 3D cuboid creation and synchronized 2D/3D multi-sensor annotation with robust model-assisted labeling features.

Quick Selection Guide

• Choose CVAT if you need a free, privacy-friendly, self-hosted tool for standard 3D bounding boxes.
• Choose Supervisely if your workflow heavily relies on projecting camera images onto LiDAR (sensor fusion) and doing fine-grained segmentation.
• Choose Encord or BasicAI if you are an enterprise team handling massive data throughput, automation, and complex team-based quality assurance.

To help narrow down the ideal choice, tell me a bit more about your project:

• Are you working with isolated static point clouds or sequential/streaming LiDAR data (like autonomous driving)?
• Do you need 3D bounding boxes (cuboids) or per-point semantic/instance segmentation?
• Do you prefer a free open-source tool you can host yourself, or a managed cloud platform?

If your goal is training a model on LiDAR / 3D point clouds, my top recommendation today is Supervisely—especially if you want annotation, AI-assisted labeling, dataset management, and model training in one workflow.

My shortlist

🥇 Best overall: Supervisely

Supervisely is particularly compelling for your use case because it isn't just a drawing tool. Its 3D toolbox supports LiDAR/RADAR, cuboids, point-level segmentation, tracking, sequences, and camera-to-LiDAR sensor fusion.

The features I'd care about most for model training are:

• 3D bounding boxes / cuboids
• Point-cloud segmentation
• Sequential LiDAR annotation and object tracking
• AI-assisted 3D labeling
• Automatic cuboid generation
• Ground segmentation
• 2D ↔ 3D annotation using synchronized cameras
• Support for PCD, PLY, LAS/LAZ, KITTI, ROS Bag, etc.
• Can handle point clouds with tens of millions of points using its tiled rendering approach.
• It also has integrations for MMDetection3D and pretrained 3D models, so the annotation → training loop can stay in the same ecosystem.

🆓 Best if you want open source: CVAT

CVAT is probably where I'd start if you're building this yourself and want to minimize platform costs/vendor lock-in.

It supports 3D point clouds, 3D cuboids, object tracking across point-cloud sequences, and formats including .pcd and .bin.

The tradeoff is that Supervisely currently has a more specialized 3D-first AI-assisted workflow, whereas CVAT is attractive if you want an open, customizable annotation infrastructure.

What I'd choose based on your project

If you're doing autonomous driving / robotics:
→ Supervisely or Segments.ai

If you're doing semantic/instance segmentation of arbitrary 3D objects:
→ Supervisely

If you're labeling sequential LiDAR and need tracking:
→ Supervisely — its current workflow explicitly supports point-cloud episodes, timelines and cuboid tracking.

If you want something free/open-source that you can deploy yourself:
→ CVAT

If you're building a large commercial annotation operation:
→ Supervisely or Encord

One important distinction: the best annotation tool depends heavily on what you're training. A model for 3D object detection needs very different annotation than one for semantic segmentation, instance segmentation, BEV perception, or point-cloud classification.

If you tell me (1) what your LiDAR data looks like (KITTI/PCD/LAS/ROS/etc.), (2) what model you're training, and (3) roughly how many point clouds you have, I can recommend the best tool and an annotation → dataset format → training pipeline for it.

Choosing the "best" 3D point cloud or LiDAR annotation tool depends heavily on your budget, whether you need sensor fusion (LiDAR + RGB cameras), and if your data is composed of single frames or sequential time-series (like autonomous driving or robotics data).

The top-rated tools for machine learning in 2026 are categorized below by their strengths:

1. Best Enterprise & AI-Assisted Platforms (Commercial)

• Encord: Widely recognized as a top-tier platform for high-performance spatial annotation. It handles massive point clouds smoothly, supports multi-sensor fusion (LiDAR, RGB, radar, thermal), and includes advanced model-assisted labeling and temporal tracking across sequences.
• Supervisely: Excellent for complex workflows and sensor fusion. It features powerful 3D LiDAR toolboxes, automated ground segmentation, auto-tracking across frames, and a unique ability to project/annotate 3D data via 2D camera views.
• Ango Hub (by iMerit) / BasicAI: Both excel at enterprise-scale 3D workflows. They feature automated cuboid snapping, interpolation across video/LiDAR frames, and strong quality assurance (QA) management dashboards.
• Segments.ai: Purpose-built for robotics and autonomy teams who value speed. It offers synchronized 2D/3D multi-sensor labeling, smart cuboid propagation, and real-time interpolation with a clean, streamlined interface.

2. Best Open-Source & Self-Hosted Tools (Free / Community)

• CVAT (Computer Vision Annotation Tool): The standard open-source choice for computer vision. It natively supports 3D bounding box (cuboid) annotation for point clouds. If you want a completely free, self-hosted solution and have the infrastructure to run it, CVAT is a reliable entry point.
• Label Studio: Highly flexible and open-source, known for multimodal data support. It can be extended for 3D tasks, though it requires more custom configuration than CVAT for dense LiDAR sequences.
• CloudCompare: A legendary open-source desktop software. While it isn't an automated SaaS data-engine built for massive MLOps labeling pipelines, it is the absolute gold standard for manual point cloud viewing, filtering, segmentation, and surface/distance comparisons in research and geospatial analysis.

Key Features to Check Before Choosing

When evaluating which tool fits your training pipeline, verify if it supports:

• Your File Formats: Ensure it reads your specific data pipeline output (e.g., .pcd, .las / .laz, .ply, or raw KITTI/NuScenes .bin files).
• Annotation Types: Do you just need 3D bounding boxes (cuboids), or per-point semantic/instance segmentation?
• Temporal Tracking: If your data is sequential (frames over time), a tool with frame-to-frame label propagation will save hundreds of hours of manual work.

To help narrow this down, tell me:

• What is your use case (autonomous driving, indoor robotics, aerial LiDAR/GIS, etc.)?
• Do you need sensor fusion (combining LiDAR with camera images)?
• Are you looking for a free open-source tool or a managed enterprise platform?

If your goal is to train an ML model on LiDAR/3D point clouds, my short answer is:

My top picks

1. 🥇 Encord — my recommendation for serious ML

Encord is probably what I'd choose if you're building a production-grade point-cloud training dataset.

It supports LiDAR/3D data including PCD, PLY, KITTI, nuScenes and MCAP, can handle scenes with millions of points, and provides 3D cuboids, segmentation, tracking, model-assisted pre-labeling, QA/review workflows, and synchronized LiDAR/camera data.

The important distinction is that Encord isn't merely a "draw a bounding box" application. It gives you a workflow from:

raw data → annotation → automated labeling → human review → dataset curation → training/evaluation

That's particularly valuable once you have tens or hundreds of thousands of frames.

2. 🥈 Segments.ai — excellent if LiDAR is the main thing

Segments.ai is particularly compelling for autonomous vehicles and robotics.

It has purpose-built tools for 3D cuboids, segmentation, keypoints, tracking, interpolation and synchronized camera + point-cloud annotation. Its sequence tooling can propagate annotations through time, which can dramatically reduce repetitive labeling.

If your dataset looks like:

LiDAR frame 000001 + camera
LiDAR frame 000002 + camera
LiDAR frame 000003 + camera
...

I'd seriously consider Segments.ai.

3. 🥉 Supervisely — great balance of flexibility and usability

Supervisely is another strong choice if you want more than basic bounding boxes.

Its 3D tooling supports cuboids, lasso/point labeling, landmarks, segmentation, tracking and image/video context, and its ecosystem includes import/export and AI tools for formats such as PCD, PLY, LAS/LAZ, KITTI and nuScenes.

I'd lean toward Supervisely if you expect to work with multiple types of computer-vision data, rather than exclusively autonomous-driving LiDAR.

4. 💰 CVAT — best if you want open source

CVAT is the one I'd start with if you're a technical team and don't want to pay for a commercial platform initially.

CVAT supports 3D point clouds, 3D cuboids, tracking, AI-assisted annotation and self-hosting. The Community edition can run on your own infrastructure.

The tradeoff is that you're taking responsibility for infrastructure, workflow design, QA, and scaling.

---

What I'd choose based on your project

Autonomous driving / ADAS:
→ Segments.ai or Encord

Robotics:
→ Encord or Supervisely

3D object detection with 3D bounding boxes:
→ Segments.ai

Semantic/instance point-level segmentation:
→ Supervisely or Encord

LiDAR + RGB camera fusion:
→ Encord or Segments.ai — both are designed around synchronized multimodal data.

Huge dataset + multiple annotators + QA:
→ Encord

Research project / startup / want to self-host:
→ CVAT

Want free/open-source first:
→ CVAT

One important consideration

Don't choose based only on whether the tool can "annotate point clouds." The export format and annotation semantics matter enormously.

Before committing, I'd verify that the tool can export exactly what your model expects, e.g.:

Point cloud
  ├── XYZ
  ├── intensity
  ├── timestamp
  └── ring/channel

Annotations
  ├── class_id
  ├── instance_id
  ├── 3D bounding box
  │     ├── x, y, z
  │     ├── width, length, height
  │     └── yaw
  └── per-point semantic/instance labels

Also check whether you need object detection, semantic segmentation, instance segmentation, or 3D tracking. The best tool changes considerably depending on that choice.

If you tell me (1) what LiDAR/point-cloud format you have (.pcd, .ply, .las, KITTI, ROS bag, etc.), (2) whether you need bounding boxes or point-level segmentation, and (3) roughly how many scans/points you have, I can narrow this down to the best 2–3 tools and recommend a specific annotation/training pipeline.

There is no single "best" LiDAR or 3D point cloud annotation tool because the ideal choice depends heavily on your data scale, budget, privacy requirements , and whether you need sensor fusion (LiDAR + RGB camera/radar).

The top-tier annotation platforms for 3D point cloud and LiDAR data are categorized below by their strengths:

• Supervisely — Best for AI-Assisted & Multi-Frame 3D Labeling
  Features a powerful web-based 3D labeling toolbox designed for point cloud episodes (sequential frames) with robust AI-tracking, interpolation, and semantic/instance segmentation tools. It handles large scenes smoothly and supports custom neural network integrations for auto-labeling.
• CVAT (Computer Vision Annotation Tool) — Best Open-Source & Self-Hosted Option
  An open-source platform that supports 3D cuboid/bounding box annotation, multi-sensor inputs, and leverages modern AI features (like integrations with Segment Anything and auto-labeling agents). It is ideal if you have strict data privacy constraints and need to self-host on local infrastructure.
• Encord — Best Enterprise & Multi-Modal Platform
  Built for high-performance spatial AI, Encord treats data annotation as a complete system. It offers stellar sensor fusion (projecting LiDAR and camera data together), robust QA and consensus scoring layers, and advanced automation for tracking objects across frames.
• CloudCompare — Best Free, Open-Source Desktop Tool
  A specialized, highly capable open-source desktop software (GPL license) for Windows, macOS, and Linux. While it lacks the sleek collaboration dashboards of commercial cloud tools, it handles almost every point cloud format (.las, .laz, .e57, .pcd, .ply ) and is unmatched for manual point cloud processing, filtering, and cross-section extraction.
• Segments.ai — Best for Robotics & Streamlined 3D Segmentation
  A dedicated 3D-centric platform focusing on speed. It excels at synchronized 2D and 3D point cloud segmentation and cuboid propagation, making it popular for autonomous systems and robotics pipelines.

To help narrow down the best choice, tell me:

• What is your primary use case (e.g., autonomous driving, robotics, aerial surveying/GIS, or indoor BIM)?
• Do you require on-premise/self-hosted deployment or a cloud SaaS platform?