Junhyeok Lee

Applies to v0.1.0.

Camera

The camera matches the module on the real car. One placed camera can produce up to three image streams: colour (RGB), depth, and segmentation. When you finish this page you will have a camera on the car, know what each setting does, and see its picture live.

Add it

Sensor Setup → Sensor Type CameraVEYE IMX462 under Sensor Selection → place it. Then open its sections below.

Intrinsics: the picture itself

FieldWhat it doesStock value
Resolution (Width × Height)Pixel size of every stream1920 × 1080
Field of View (Horizontal / Vertical / Diagonal)How wide the camera sees, in degrees. The three values are linked through the aspect ratio — set one, the others follow.matches the real module's ~51° horizontal
Frame RateFrames per second on the wire30 Hz

Noise: make the colour image imperfect on purpose

A real image sensor is not clean. This section adds that imperfection to the RGB stream so what you train on is not suspiciously perfect:

  • Enabled — master switch. Off means a clean render, whatever the values below say.
  • Model — which noise formula to apply.
  • Read σ — constant sensor noise, the same in dark and bright areas.
  • Shot σ — brightness-dependent noise, stronger in bright areas.

σ (sigma) is the noise strength: 0 is none, bigger is noisier. Leave this off while you are learning the simulator; turn it on when you start training and want robustness. See Sim-to-Real.

Depth Camera: distance per pixel

  • Enable — turns the depth stream on. Off by default.
  • Topic — the depth stream's own topic name (read-only).
  • Max Range (m) — the farthest distance the depth image encodes.
  • Noise Model, σ₀, σ_q — depth noise: σ₀ is a constant floor, σ_q grows with distance, matching how real depth sensors get worse farther out.

Segmentation: the label image

  • Enable — turns the segmentation stream on. Off by default.
  • Topic — its topic name (read-only).

Every pixel is painted the single flat colour of its class (road, lane, obstacle …) with no smoothing, so the image doubles as an exact ground-truth label. This is the stream the dataset collector records.

See it working

  • Foxglove — an Image panel on the camera's topic.
  • ROS — the streams are compressed images; the Visualisationpage shows the three ways to view them (rviz's Image display alone will not; its dropdown only lists raw images).

Common problems

  • The dataset collector says no segmentation image — the Segmentation stream is off. Enable it here.
  • Depth/segmentation topics missing from ros2 topic list — same cause; a stream that is off publishes nothing.