Semantics

The elevation map can also include semantic information. On the ROS2 branch the workflow consists of two pieces:

semantic fusion inside elevation_mapping_cupy
semantic inputs produced by the in-repo semantic_sensor package or an external vision model

Semantic Fusion in the Elevation Map

Semantic inputs are configured in the robot sensor YAML under the subscribers section. Each subscriber can define channel names and a fusion mode per channel.

The channel list maps incoming semantic channels to map layers. The fusion list defines how each channel is fused into the persistent map.

Supported fusion modes include:

average

Computes a weighted average between the current cell value and the newly observed value.

Use case: dense semantic features

bayesian_inference

Applies Gaussian Bayesian fusion, using the previous posterior as the next prior.

Use case: continuous semantic features

class_average

Averages per-cell class scores while ignoring uninitialized cells.

Use case: class probabilities

class_bayesian

Applies Bayesian fusion on a categorical distribution with a Dirichlet prior.

Use case: class probabilities

color

Fuses packed RGB color channels.

Use case: RGB appearance layers

Semantic Input Producers

Sensors do not always publish semantic channels directly. The in-repo semantic_sensor package provides two launchable producers:

semantic_pointcloud.launch.py generates a multi-channel pointcloud.
semantic_image.launch.py generates semantic image channels plus ChannelInfo metadata.

The supported end-to-end ROS2 examples are:

turtlesim_semantic_pointcloud_example.launch.py
turtlesim_semantic_image_example.launch.py

The intended architecture is simple: a vision model produces semantic channels, and elevation_mapping_cupy fuses those channels into map layers.