Fundamentals of 3D LiDAR annotation at scale

A LiDAR label is what your perception model trains on, so it has to carry more than a box. In a production dataset each object gets a 3D cuboid or a per-point mask, a class, a tracking identity, a motion state and a situational reasoning attribute. Sycrad’s new export system, described on the roadmap, is designed to carry these labels straight into model training with no pre-processing.

Point clouds are worth the effort because they give you metric depth. Every point has real x, y, z coordinates and a return intensity, so you measure an object’s size, heading and trajectory instead of inferring them from pixels. What your sweeps look like depends on the sensor. Spinning sensors cover a full 360° and are dense nearby but thin at range. Solid-state and MEMS sensors cover a narrower cone at higher angular resolution, and FMCW sensors add per-point radial velocity through the Doppler effect.

The 3D bounding cuboid is the standard annotation primitive for object detection and tracking. Each 9-DOF cuboid defines an object’s center, dimensions and orientation in a single label. Sycrad currently supports 3D cuboid annotation, while panoptic segmentation, which pairs per-point semantic classes with instance IDs for individual objects, is on the roadmap.

Sycrad is designed for the four domains where LiDAR ground truth matters most: autonomous vehicles, robotics, defense and aerial mapping. Its local-first architecture keeps sensor data on your own infrastructure. Broader dataset coverage across all four is part of the roadmap.

Volume is what breaks manual labeling. A single sequence runs to hundreds or thousands of sweeps, and placing cuboids on every one of them by hand does not scale. Sycrad addresses this with its fundamental features, interpolation between two anchor frames, and SLAM pre-labeling across the whole sequence. Together they substantially reduce the manual labeling load without any AI model. Planned AI-assisted labeling will support both local ONNX inference and cloud-hosted models, with details on the roadmap.

Core annotation workflows

Manual frame-by-frame labeling is unsustainable for large sequences. Sycrad offers three workflows, from single-frame precision to automated multi-frame pipelines:

Single-frame annotation workflow: Each sweep can be inspected and annotated individually. Place one-click oriented cuboids with Auto-Fit (A key) and check them in top, front and side views, while the camera channel follows your viewport and shows the image that matches the direction you are looking. Cuboids project directly onto the camera images, so you can verify each box against what the camera sees. Each object is then assigned a persistent tracking ID, and the Tracking ID Manager tracks and controls cuboid states across frames, leveraging the interpolation core so annotators can advance boxes to subsequent sweeps with a single action. These core capabilities serve as the shared foundation across both interpolation and SLAM workflows.
Anchor-based interpolation workflow: Annotators place bounding cuboids at two distinct anchor frames. Sycrad's interpolation engine registers consecutive sweeps and resolves the intermediate trajectory, correcting kinematic drift and filling in-between frames with consistent bounds. Learn more in our interpolation workflow guide.
SLAM-based pre-labeling workflow: Instead of labeling individual sweeps in isolation, Sycrad’s SLAM pipeline merges consecutive scans into a dense, high-contrast composite scene. The system separates dynamic points from static background elements, clusters candidate moving objects, and generates continuous trajectories across the full sequence. Static objects are placed once and mapped automatically to the frames where they are visible, turning hours of repetitive labeling into a fast verification pass. Learn more in our SLAM workflow guide.

Supported dataset formats

KITTI Raw, nuScenes v1.0, and Waymo Open Dataset (v2 Parquet): Fully supported across every Sycrad capability, including interpolation, SLAM pre-labeling, Auto-Fit, the Tracking ID Manager, import and export, and synchronized camera projection. Waymo sequences are ingested directly via an embedded Apache Parquet reader and converted into optimized binary sweeps inside Sycrad without offline conversion bottlenecks.
PCD & LAS/LAZ: Industry-standard point cloud formats currently supported for inspection and visualization. Annotation workflows are under active development.

Release history and completed milestones are documented in the changelog.

Solving the 5 bottlenecks of point cloud labeling

Annotating 3D point data introduces fundamental spatial challenges that differ sharply from 2D pixel labeling:

Sparsity at distance: Objects located far from the sensor often contain only 3 to 10 points, making precise boundary determination difficult. Sycrad addresses this with SLAM multi-frame accumulation, integrated sensor fusion that automatically matches the viewport to the correct calibrated camera channel, and 2D cuboid projection onto camera images. Planned Gaussian Splatting support will add photo-realistic 3D scene reconstruction to help verify sparse returns.
Occlusion and perspective shadows: Laser beams cannot penetrate solid obstacles, leaving void regions behind structures. Sycrad addresses this by merging temporal sweeps through SLAM and overlaying multi-camera imagery, allowing annotators to reconstruct complete object boundaries from complementary perspectives.
Dynamic object motion: Moving actors smear into fragmented trails in raw multi-frame overlays. Sycrad’s SLAM pipeline isolates these dynamic returns and consolidates them into a unified object frame, sharpening object geometry so moving actors can be reviewed and labeled in a single click.
Weather artifacts: Rain, fog, dust, and exhaust plumes reflect laser pulses, generating stray spatial noise. Adaptive weather filtering that suppresses atmospheric reflections and isolates clean ground truth is currently in development.
Scale: LiDAR sequences span hundreds or thousands of sweeps, and merging them into a single SLAM scene or loading a large aerial scan multiplies the point count far beyond any single frame. While containerized and browser-based WebGL tools stutter under heavy data loads, Sycrad’s custom Qt QRhi graphics pipeline communicates directly with native graphics APIs to process massive point sequences smoothly on local workstations.

Air-gapped data sovereignty

LiDAR point clouds frequently record proprietary, classified, or regulated environments. Defense programs operate under strict security clearances where sensor feeds cannot touch third-party infrastructure, and regulations such as ITAR strictly prohibit transferring defense spatial data to external cloud servers. Automotive manufacturers treat unreleased vehicle test runs as trade secrets, and public roadway recordings fall under GDPR because they can contain personally identifiable information. Robotics teams deploying in customer warehouses cannot upload facility layouts or inventory positions to third-party cloud servers without violating SOC 2 mandates.

Some enterprise platforms offer self-hosted Docker deployments to reduce external data transfer, but they are often limited by browser-based rendering and container memory overhead. Many existing solutions are also built around external GPU clusters or remote AI APIs, which makes it hard to run standalone algorithmic pipelines or lightweight local inference on a standard workstation without dedicated backend infrastructure.

By contrast, Sycrad is a self-contained, local-first desktop engine. Raw point cloud data stays on your workstation storage and never leaves your machine. For organizations with strict air-gap requirements, Sycrad Enterprise runs entirely offline, with zero external network calls, zero telemetry and on-premise license validation.

Why choose Sycrad?

When selecting an annotation solution for autonomous vehicles, robotics, defense, or aerial mapping, engineering teams typically weigh four criteria:

Data Sovereignty: Cloud platforms require uploading large datasets to third-party servers, which adds ongoing bandwidth and egress costs. Sycrad keeps all point cloud data on local storage.
Performance: Browser apps run into JavaScript garbage collection pauses and WebGL memory limits. Native workstation software talks to the operating system’s graphics APIs directly, which keeps viewport interaction fluid and low-latency.
Dataset Compatibility: Sycrad natively supports autonomous vehicle datasets for annotation and geospatial point cloud formats for inspection, with no lossy conversion steps. Broader format coverage will continue to expand, and Enterprise clients can request custom format integrations tailored to proprietary sensor configurations.
Annotation speed: Throughput is the primary bottleneck in spatial perception data engines. Sycrad addresses this with deterministic algorithms, pairing SLAM pre-labeling, Auto-Fit, and the Tracking ID Manager with planned local and cloud AI-assisted features. Teams will be able to choose between fully private on-device ONNX inference and optional cloud acceleration. This hybrid approach lets teams produce high-accuracy ground truth across large point cloud datasets in substantially less time and at a fraction of traditional operational costs.

For engineering teams working across autonomous vehicles, robotics, defense and aerial mapping, Sycrad delivers workstation-class speed and complete data confidentiality.