deepstream-run-mv3dt

Resolve the deepstream-tracker-3d-multi-view app directory, verify prerequisites with the setup script's check-only mode, then run the 4-camera sample by default. Display mode shows OSD and BEV windows; headless mode saves a tiled MP4 and a Kafka-derived BEV MP4.

Place videos under a dataset's videos directory with matching camInfo calibration files, then generate configs with the auto-configurator and launch DeepStream in Docker. If calibration is missing, the workflow delegates to AutoMagicCalib and normalizes modelInfo for your chosen detector.

MV3DT supports PeopleNetTransformer as the default people detector, RTDETR as a warehouse multi-class detector, and PeopleNet v2.6.3. The selected DETECTOR_MODEL drives both config generation and calibration modelInfo normalization.

Yes, headless mode generates configs with file output and message broker enabled, producing a tiled DeepStream MP4 and an offline BEV MP4 captured from Kafka. No X11 display is required for this path.

Empty BEV output usually means the offline capture started too early or too late relative to the DeepStream launch, or Kafka offsets never advanced. Start a fresh capture with --from-end immediately before the Docker launch and verify nonzero message and frame counts in the capture log.

The reference launch uses those flags for broad device, network, and mounted-repo access, so the Skill requires explicit user approval before executing. It also documents least-privilege alternatives such as --gpus all, specific device mounts, and explicit port mappings.