3D Occupancy Grid Annotation for Autonomous AI

3D occupancy grid annotation divides a 3D scene into cubic voxels (volumetric pixels) and classifies each of them as occupied, free, or unknown. This approach is used for training unmanned vehicles because it allows modeling arbitrary geometric shapes and objects that are difficult to describe with standard 3D bounding boxes.

Quick Take

Traditional 3D bounding boxes capture too much empty space around irregular shapes.
Voxel annotation divides environments into precise 3D grids.
Advanced datasets predict the trajectories of hidden obstacles using context-aware occluded object estimation.
BEV format streamlines trajectory math.
Human annotators remain essential to fix automated grid defects like "ghost" voxels.

Solving the Geometry Problem for Autonomous AI

Traditional rectangular 3D frames work well for simple, blocky shapes on the road, such as standard sedans or shipping containers.

However, autonomous vehicles frequently encounter complex, irregular geometries that rectangular boxes cannot measure accurately. When a bounding box encloses a spilling load of gravel, a low-hanging tree branch, or a deformed vehicle body after a crash, the box inevitably includes a large volume of empty air, which confuses the self-driving AI.

To solve this spatial inaccuracy, voxel annotation divides the entire 3D environment into a grid of tiny, equal-sized cubes. Instead of drawing a single loose box around an object, annotators label each individual voxel as either occupied or empty. This grid-based approach ensures that the AI learns the exact physical boundaries of complex shapes, preventing dangerous navigation mistakes caused by empty-space miscalculations.

Why Standard 3D Boxes Fail on Irregular Objects

Three main limitations make 3D boxes highly problematic for modern self-driving systems:

Excessive Empty Space. A bounding box around a diagonal barrier or a tree branch labels the surrounding empty air as a solid obstacle, forcing the vehicle to make unnecessary, sudden stops.
Inability to Map Details. Fine-grained elements, like an open truck door or a pedestrian's extended arm, get completely lost inside a large, generic box.
Poor Occlusion Handling. Traditional boxes cannot reliably show which parts of a vehicle are hidden behind a wall, making accurate occluded object estimation nearly impossible for prediction algorithms.

Using voxel annotation bypasses these limitations entirely by shifting the focus from "identifying objects with boxes" to "mapping every cubic centimeter of physical space".

Technical Comparison

While bounding boxes require less storage space and are faster to draw, they lack the high-fidelity spatial awareness required for safe urban navigation.

The following table compares structural differences in data metrics between standard 3D bounding boxes and occupancy grid labeling.

Evaluation Metric	3D Bounding Boxes	Occupancy Grid Labeling
Handling Irregular Objects	Poor: boxes capture too much empty space around debris, cranes, or foliage.	Excellent: labels the exact physical shape of any object voxel-by-voxel.
Dataset Storage Size	Small: only requires 3D coordinates, orientation, and size parameters.	Large: requires dense 3D arrays to store the state of millions of individual voxels.
Free-Space Accuracy	Low: often misidentifies drivable, empty spaces as blocked zones.	High: clearly defines completely empty paths for safe vehicle routing.
Coordinate-Level Semantics	None: class labels are applied to the whole box, not individual points.	Complete: every single 3D coordinate has an assigned semantic class (e.g., road, pole, car).

Using occupancy grid labeling ensures that the vehicle's planner receives a highly detailed 3D map, allowing the path-planning algorithm to find safe gaps and drivable paths that a box-based system would simply label as completely blocked.

Safe Path Planning

Mapping Navigable Zones via Free-Space Annotation

Safe path planning requires the system to have a clear understanding of the physical boundaries of the roadway. Free-space annotation technology addresses this task by marking the boundaries of zones suitable for movement. Instead of simply ignoring unoccupied areas, annotators verify every centimeter of the clear path, creating a reliable map of safe space.

This approach ensures proper autonomous vehicle behavior in complex urban scenarios where standard detection methods fall short:

Driving under low clearance objects – accurately defining the free space under arches, bridges, or tree branches, where the vehicle's vertical clearance needs to be calculated.
Maneuvering in tight spaces – driving in narrow driveways between buildings, where the road boundaries run right up against curbs, utility poles, or parked cars.
Temporary road changes – navigating past construction sites with non-standard barriers and cones that restrict the usual traffic lane.

Predicting Trajectories with Occluded Object Estimation

The greatest threat to autonomous vehicle safety comes from "blind spots" – areas of the road hidden from sensors by other vehicles, buildings, or thick fog. When a large truck blocks the camera or LiDAR line of sight, the perception system cannot see what is happening directly behind it.

To solve the occlusion problem, modern datasets integrate occluded object estimation algorithms. This labeling method allows for marking space while accounting for the probabilistic occupancy of hidden areas.

Based on the analysis of preceding frames and the general context of the scene, this markup helps the AI model predict the trajectory of hidden objects (for example, a pedestrian stepping behind a bus). Thanks to this, the autonomous vehicle gains the ability to anticipate potential danger even before the object physically appears in the sensors' direct line of sight, allowing it to reduce speed in a timely and smooth manner.

Scaling BEV Occupancy Dataset Creation

The specificity of the BEV format is that it translates complex three-dimensional sensor data into a flat two-dimensional projection from above. This allows the drone to assess the situation around it 360 degrees in a single coordinate system, greatly simplifying trajectory calculation.

Automatic point cloud conversion

Our standard pipeline begins with the automatic conversion of raw LiDAR point clouds and camera images into a preliminary 3D voxel grid, which is then projected into the BEV format. However, automation is not perfect, and the initial processing always contains defects that the AI model cannot correct on its own.

The most common automation defects are:

"Ghost" voxels. The algorithm mistakenly marks noise from dust, exhaust gases, or fog as solid obstacles on the BEV map.
Blurred boundaries. The merging of real object voxels with the surrounding space or road surface.
Loss of thin geometries. Omitting voxels that represent wires, road signs, or thin metal fences.

Without manual validation, these errors can end up in the training data, directly reducing the driving safety of the autonomous vehicle. Thus, to eliminate automation errors, Keymakr annotators manually review and correct each scene within the Keylabs interface. Special attention is paid to complex cases – fast-moving dynamic objects and occlusion zones. Our specialists manually restore object geometry in the BEV projection, separating real obstacles from visual noise.

FAQ

What is the main difference between semantic voxel markup and ordinary 3D bounding boxes?

3D bounding boxes mark an object as a single whole, whereas voxel annotation divides the entire space into tiny cubes and classifies each of them individually. This allows the autonomous vehicle to see the exact physical shape of the object, fine details, and deformations, which is critical for safe maneuvering in tight spaces.

How does a BEV occupancy dataset help save the autonomous vehicle's computing resources?

A full three-dimensional voxel grid creates a colossal load on the vehicle's onboard computer. The BEV format compresses this 3D information into a flat 2D top-down projection. The resulting BEV occupancy dataset allows the AI model to quickly analyze free and occupied zones around the car, consuming several times less memory and processor time.

How do annotators process dynamic objects in 3D occupancy network data?

Moving objects are marked taking into account the temporal factor. Annotators use specialized tools that allow them to track frame sequences. This guarantees that voxels belonging to a moving car or pedestrian are correctly transferred from frame to frame, without leaving "ghost" trails in the final data grid.

To plan a safe path, it is not enough for the AI to only know where the obstacles are. Using free-space annotation, markers clearly outline the boundaries of guaranteed free and safe space. This prevents situations where the car brakes sharply in front of safe objects or cannot drive under a low arch due to an incorrect height evaluation.

When one object occludes another, the autonomous vehicle's sensors lose visibility. Annotators mark hidden zones as "likely occupied" based on the scene's context. This trains the model to anticipate danger and smoothly reduce speed in advance before closed sections of the road.