LiDAR annotation without the tool sprawl: point clouds, camera fusion, and tracking in one workspace

Most perception teams annotate the same scene four times. Your annotators draw cuboids (3D bounding boxes) in a point cloud tool, check camera frames in a second app, track objects across frames in a third, then export everything to a spreadsheet for review. Every handoff drops context, and every dropped context becomes rework two weeks later when a reviewer finds a cuboid that drifts between frames.

We built a LiDAR and multi-sensor annotation interface that keeps the whole scene in one place. It runs as an Interface inside Label Studio Enterprise, so your team labels point clouds, reviews calibrated camera views, follows objects through time, and gates quality without leaving the task. Here is how it works.

Track objects through time without re-drawing every frame

A car moves through a scene while its cuboid needs to follow. Set a cuboid on a keyframe, set another a few frames later, and the interface interpolates the positions in between. When your annotator nudges a box on one frame, the surrounding frames re-interpolate right away, so following an object from the frame it enters to the frame it leaves takes a handful of edits instead of one per frame. Fewer manual frames means fewer identity swaps, and fewer identity swaps means fewer silent errors in your training set. This is helpful whether you're building for autonomous driving, ADAS, or robotics.

Sensor fusion: see every camera and LiDAR view from the right angle

Select a cuboid and the camera panel switches to the view that actually contains it, so your annotator confirms a hard object against the footage instead of guessing from the point cloud. They choose which LiDAR sensors to render, color the points to expose structure, and split the workspace to compare views side by side. A 2D top-down view gives them the whole scene at once when the 3D view gets crowded.

Start from predictions, keep human control

Import your LiDAR data with model predictions attached, and the cuboids show up in the scene as suggestions. Your annotator selects one, accepts it, and it becomes a real cuboid they can adjust: reposition it, resize it, correct its heading. The model does the first pass, and the person decides what stands.

While your annotator draws, the interface highlights the LiDAR points that fall inside the cuboid, so they can see whether the box actually captures the object or clips half a car. The geometry goes beyond cuboids: points, lines, lanes, and point cloud segmentation with semantic and instance masks, so a lane boundary and a pedestrian each get the shape they need. Lidar segmentation runs in the same view as cuboid tracking, so your annotator switches shapes without switching tools. Object IDs stay stable across the sequence, and provenance travels with each object, so a reviewer always knows where an annotation came from.

Handle real-world data sizes

Production LiDAR files get large. The interface reads COPC files with more than 10 million points and loads them on demand: as your annotator zooms and navigates, it pulls in the detail for the area they are working in. A detail budget lets them load more resolution with one click when they need it, so a heavy scene stays responsive.

Validate before delivery

Quality checks live inside the task rather than in a separate audit pass. Reviewers move through validation view by view, blocking errors stop a task from being marked done, and undo and recovery let someone fix a mistake without restarting the scene. You catch the bad annotation before it reaches the model, not after.

What this changes for your team

Keeping annotation, sensor fusion, temporal tracking, and QA in one interface does three things: your team iterates faster, redoes less work, and delivers training data your ML engineers can trust.

The LiDAR workspace is one HumanSignal Interface. The same system lets your team build annotation screens shaped to whatever modality you work in, from documents to audio to the multi-sensor scenes above.