MV3DT Vision AI Profile

Overview

Introduction

The VSS Warehouse Blueprint’s MV3DT Vision AI Profile is a comprehensive guide to building a 3D intelligent video analytics system using RT-DETR object detection and Multi-View 3D Tracking (MV3DT). Unlike the 3D Vision AI Profile, which uses the Sparse4D end-to-end 3D perception model, this profile pipelines a 2D detector (RT-DETR) with a distributed multi-view 3D tracking framework (MV3DT) to produce 3D Bird’s Eye View (BEV) outputs from multiple overlapping cameras. It provides a detailed overview of the system architecture, data flow, and key components.

Deployment Architecture

Components and Interactions

The VSS Warehouse MV3DT Blueprint emphasizes 3D multi-camera detection, tracking, and behavior analytics for safety events and metrics using RT-DETR and MV3DT. Below is a breakdown of the components and their interactions.

1. Input Source

   • Videos: Raw video data stored in a filesystem, serving as input for processing.
   • NvStreamer (NVStreamer): A microservice that streams videos via RTSP (Real-Time Streaming Protocol) to the VIOS (Video IO & Storage). NvStreamer can be swapped with real-world cameras.

2. Video IO & Storage (VIOS) (VIOS Microservices)

   • VIOS ingests video streams from NvStreamer via RTSP.
   • It records the streams and forwards them (via RTSP) to the DeepStream microservice for further processing.

3. DeepStream (RT-DETR + MV3DT) (3D Multi Camera Detection and Tracking (MV3DT))

   • DeepStream processes RTSP streams for 3D multi-camera detection and tracking, combining the RT-DETR detector (RT-DETR (2D warehouse detection)) with the MV3DT distributed tracking module.
   • RT-DETR generates 2D bounding boxes for warehouse-relevant object classes. The MV3DT module back-projects these 2D detections into 3D world coordinates using each camera’s projection matrix and 3D object models, and fuses measurements across cameras with overlapping fields of view.
   • Cameras exchange 3D tracklets with their vision neighbors through an MQTT Broker for decentralized ID negotiation, cross-camera ID propagation, and multi-view measurement fusion.
   • It sends per-camera frame data, including detected and tracked object IDs with 3D foot locations, in Protobuf format to the message broker via the mdx-raw topic. When appearance ReID is on, those payloads also carry CLIP-ReID embeddings.
   • The tracker queries the ReID service over HTTP (reid-embed:8088) so Person tracks can be re-associated from the appearance gallery.

4. MQTT Broker (Mosquitto)

   • An Eclipse Mosquitto MQTT message broker used exclusively for inter-camera MV3DT communication.
   • Each camera publishes its 3D tracklets to a dedicated MQTT topic and subscribes to topics from its vision neighbors (cameras with overlapping fields of view).
   • This enables decentralized, globally unique ID assignment and cross-camera handover without a central coordinator.

5. Appearance ReID (reid-embed + Milvus) (3D Multi Camera Detection and Tracking (MV3DT))

   • reid-embed consumes mdx-raw, stores CLIP-ReID (1280-D) features in a Milvus gallery (etcd + MinIO), and answers HTTP queries from the tracker. This helps re-associate people across gaps in camera coverage or when they reappear after an extended period. Re-association is Person only.
   • Warehouse deployments also run a secondary SigLIP 2 encoder (siglip_v2_vdeployable_v2.0) and publish mdx-compressed-embeddings on the same Kafka or Redis broker. SigLIP runs only on selected high-quality samples of each object, at regular intervals, so each object still gets a strong appearance embedding without putting SigLIP on every DeepStream frame. BEV Fusion still consumes only mdx-raw.
   • Gallery backends are ephemeral. CLIP-ReID and SigLIP 2 weights persist under $VSS_DATA_DIR/models/reid (Compose) or the RT-CV models PVC (Helm). Perception waits on GET /health/ready after the init container and Milvus are up. The health-check start grace period is 300s; first-boot engine build can take longer.

6. BEV Fusion (3D Multi Camera Detection and Tracking (MV3DT))

   • This microservice consumes per-camera mdx-raw data (Protobuf) from the message broker.
   • It fuses 3D measurements from multiple cameras into unified Bird’s Eye View (BEV) frames.
   • The resulting fused BEV data is published to the message broker via the mdx-bev topic for downstream consumption.

7. Message Broker (Kafka or Redis)

   • The message broker serves as the central hub for data distribution, using Protobuf for all data exchanges.
   • Kafka (Kafka): High-throughput message broker optimized for datacenter deployments with robust persistence and scalability.
   • Redis Streams: Lightweight message broker ideal for edge deployments with minimal memory footprint and low-latency requirements.
   • It also functions as a control bus, managing notifications (in JSON, via mdx-notification) for calibration updates, such as new ROI or tripwire definitions.

8. Behavior Analytics (Behavior Analytics)

   • This microservice consumes mdx-bev data (Protobuf) from the message broker.
   • It processes the data to generate behavior analytics, safety insights, and metrics.
   • The resulting data, in Protobuf format, is sent back to the message broker for indexing into Elasticsearch.

9. Storage

   • ELK (Elasticsearch, Logstash, Kibana) (ELK): Logstash retrieves BEV outputs, safety violation frames, and (when ReID secondary embedding is on) compressed appearance embeddings from the message broker, converts Protobuf to JSON, stores the data in Elasticsearch, and supports querying and visualization.

10. Visualization

• Kibana UI (Kibana UI): A user interface for visualizing analytics data stored in Elasticsearch.
• VIOS UI (VIOS UI): A separate interface for interacting with the VIOS system, receiving JSON notifications from the message broker.

11. External Interfaces

    • API Gateway and MCP (API Gateway / MCP): Enables external systems to interact with the events data through API calls.

Key Technologies

• Microservices: Components like NvStreamer, VIOS, DeepStream, Reid Embed, BEV Fusion, and Behavior Analytics are modular microservices.
• RT-DETR: Real-Time Detection Transformer for 2D object detection, pretrained on warehouse datasets.
• MV3DT: Multi-View 3D Tracking framework for distributed cross-camera 3D tracking using camera projection matrices and object models.
• MQTT: Lightweight publish/subscribe protocol for inter-camera tracklet exchange.
• RTSP: Facilitates real-time video streaming.
• Protobuf: Ensures efficient, compact data exchange.
• Message Broker: Manages data distribution and control messaging.
• ELK Stack: Supports storage, logging, and visualization.
• JSON: Used for notifications and calibration data.

Setup and Configuration

For detailed steps, see the Quickstart.

Agent Skills walkthrough

This section walks through the end-to-end MV3DT/RTVI-CV-3D journey using a coding agent (Claude Code, Codex, or NemoClaw) and the VSS Agent Skills instead. For each step it documents the example prompt you give the agent and what the agent does in response.

Use this walkthrough when you want to deploy and operate only the RTVI-CV-3D / MV3DT stack without the full Warehouse agent, LLM, or VLM stack. For the full Warehouse Blueprint, use the vss-build-vision-ai skill as described in the Warehouse Quickstart Skill section.

The MV3DT journey exercises two skills:

The agent may pause for user input during the journey. Common prompts include NGC_CLI_API_KEY for NGC image and resource access, profile size (extended for VST overlays or minimal for a smaller stack), dataset slug, video or RTSP locations, detector choice (resnet or transformer), and confirmation before starting containers or clearing state. For custom inputs, replace <path/to/files> with the directory that contains the synchronized camera MP4s, replace <rtsp-url-1>, <rtsp-url-2>, … with the ordered RTSP stream list, and replace <dataset-slug> with a short name for the camera set. Calibration flows may also ask for a Hugging Face token to stage the VGGT calibration model, or direct you to the Auto Calibration UI when layout files, alignment data, or alignment points are missing, or when you want to edit calibration parameters, ROI/tripwire definitions, or MV3DT group / region metadata.

Deploy the bundled sample dataset

Example prompts:

1. Enable multi-camera tracking on the sample dataset.
2. Deploy the standalone MV3DT stack with the bundled sample dataset.

What the agent does:

1. Loads the vss-deploy-detection-tracking-3d skill and selects the sample path, which uses the bundled warehouse-4cams-20mx20m-synthetic dataset and its in-tree calibration.
2. Chooses the extended profile by default so Elasticsearch, Logstash, Kibana, vss-video-analytics-api, and the calibration importer are deployed for VST video wall overlay rendering.
3. Detects the VSS repository path, GPU hardware profile, VSS_APPS_DIR, VSS_DATA_DIR, HOST_IP, and NGC_CLI_API_KEY.
4. Validates app-data, model, calibration, and data-log paths before starting containers. If host data-log permissions need to change, it asks before applying scoped ACLs.
5. Writes the MV3DT environment recipe (for example MODE=mv3dt, BP_PROFILE=bp_wh_kafka, STREAM_TYPE=kafka, LLM_MODE=none, and VLM_MODE=none), runs a compose dry run, and summarizes the deployment plan.
6. Waits for your confirmation before starting containers.
7. Deploys the stack and waits for readiness gates across vss-rtvi-cv-mv3dt, vss-rtvi-cv-bev-fusion, the broker, VST, configurator, behavior analytics, and the overlay services.

Deploy custom synchronized videos

Example prompts:

1. Enable multi-camera tracking on local video files at <path/to/files>.
2. Deploy standalone MV3DT with local videos at <path/to/files>.

What the agent does:

1. Loads the vss-deploy-detection-tracking-3d skill and selects the videos path.
2. Asks for any missing inputs: video directory (<path/to/files>), dataset slug (<dataset-slug>), extended versus minimal profile, detector choice, NGC_CLI_API_KEY, and whether to run independent VGGT calibration when calibration must be generated.
3. Checks whether calibration.json and matching camInfo/*.yml or camInfo/*.yaml files already exist for the requested dataset slug under the MV3DT calibration mount path.
4. If calibration is present, validates camera count, updates MV3DT camera configuration, and proceeds to deployment.
5. If calibration is missing, invokes vss-generate-video-calibration to deploy Auto Calibration, create a project from the synchronized videos at <path/to/files>, run AMC with the requested detector, optionally stage and run VGGT independently using a Hugging Face token, and export the selected calibration as MV3DT camInfo files and calibration.json.
6. If layout or alignment data is missing, pauses and directs you to the Auto Calibration UI to upload layout.png / alignment_data.json or mark and save correspondence points before calibration continues.
7. Lands the exported calibration under warehouse-mv3dt-app/calibration/sample-data/<dataset-slug>/ and makes the files readable by the containers.
8. Deploys the MV3DT stack using the custom dataset and verifies that the configured stream count matches the generated calibration.

Deploy live RTSP streams

Example prompts:

1. Enable multi-camera tracking on RTSP streams <rtsp-url-1>, <rtsp-url-2>, ...
2. Deploy standalone MV3DT against RTSP streams <rtsp-url-1>, <rtsp-url-2>, <rtsp-url-3>, <rtsp-url-4> and generate calibration first.

What the agent does:

1. Loads the vss-deploy-detection-tracking-3d skill and selects the rtsp path.
2. Collects the ordered RTSP URL list (<rtsp-url-1>, <rtsp-url-2>, …), dataset slug (<dataset-slug>), detector choice, profile size, and any user-supplied calibration path. It also checks for NGC_CLI_API_KEY and asks for a VIOS endpoint if Auto Calibration needs VIOS to capture the streams.
3. If calibration is missing, invokes vss-generate-video-calibration in RTSP mode, verifies VIOS reachability for Auto Calibration capture, creates the Auto Calibration project, runs calibration, and exports MV3DT-ready files.
4. If VGGT calibration is also requested, stages the VGGT model with a Hugging Face token and runs it independently. If VGGT does not complete, the independently generated AMC result remains available.
5. If calibration settings, layout files, alignment data, or alignment points are missing, pauses and directs you to the Auto Calibration UI to provide them before calibration continues.
6. Generates or updates MV3DT camera configuration from the calibration file, including stream count and per-camera camInfo files.
7. Deploys the MV3DT stack against the live streams and verifies that VST sensors are registered and online.

Verify and view the deployment

Example prompts:

1. Verify the MV3DT deployment and give me the VST and Kibana URLs.
2. Check whether mdx-raw and mdx-bev are flowing.

What the agent does:

1. Checks container health for perception, BEV Fusion, mosquitto, Kafka or Redis, VST, configurator, behavior analytics, and the extended overlay services when enabled.
2. Reads vss-rtvi-cv-mv3dt logs for TensorRT engine build status, active sources, and per-camera FPS.
3. Checks the BEV Fusion health status and confirms the mdx-raw and mdx-bev broker offsets are growing.
4. Queries the VST sensor list and verifies that registered sensors match the calibration camera set and are online.
5. Returns the browser URLs for VST (http://<HOST_IP>:30888/vst) and, when the extended profile is deployed, Kibana (http://<HOST_IP>:7777/kibana).

Tear down or redeploy MV3DT

Example prompts:

1. Tear down the standalone MV3DT deployment.
2. Stop MV3DT but keep calibration outputs for the next run.

What the agent does:

1. Loads the vss-deploy-detection-tracking-3d teardown flow.
2. Explains whether the request requires a plain docker compose down or a reset with down -v. For camera-set or dataset changes, it recommends resetting named volumes so stale VST sensor records and broker state do not break the next deployment.
3. Asks for confirmation before destructive cleanup such as removing named volumes, clearing data_log, or deleting Auto Calibration project state.
4. Preserves calibration outputs under warehouse-mv3dt-app/calibration/sample-data/<dataset-slug>/ unless you explicitly ask to remove them.
5. Reports the remaining preserved state and the next redeploy command path.

Testing and Validation

Kibana UI

Check for BEV, Frames and Behavior Data Views in Kibana:

1. Launch Chrome browser
2. In the address bar enter http://<IP_address>:7777/kibana

In the user interface, navigate to the Management -> Stack Management section and select Data Views under Kibana. If the data views are not visible, create a new data view (via “Create data view” button on the top right corner) for mdx-bev, mdx-frames, mdx-behavior, and mdx-compressed-embeddings-*.

Browse the Kibana UI, discover the data views and visualize the data.

VIOS UI

1. Launch Chrome browser.
2. In the address bar enter http://<IP_address>:30888/vst/.

Configure

Use the VIOS UI to configure camera streams and verify that video feeds are active.

View VIOS Video Wall

Enable Overlay settings (refer to Playback Overlay on Video Wall) to view 3D object detection and tracking results.

Events and Incidents

Events

The VSS Warehouse MV3DT Blueprint performs multi-view 3D detection and tracking (MV3DT) across calibrated multi-camera feeds, tracking objects in 3D world coordinates, and generates events based on configurable logic.

One type of event is supported:

• ROI Events: Triggered when objects enter or exit a predefined region of interest (ROI).

For more details, see Event Detection. All event data is stored in Elasticsearch, and can be queried and visualized in Kibana and the VIOS UI.

Incidents

The MV3DT profile does not generate confirmed incidents. The per-frame violation data (such as proximity and zone analytics) is still computed and written to the enhanced frames, so downstream consumers can read it directly from the message stream or Elasticsearch—but it is not promoted to tracked incidents. For the enhanced-frame violation fields, see Enhanced Frame Data.

Use Cases

ROI events provide the foundation for occupancy monitoring and safe-zone enforcement across calibrated multi-camera 3D scenes—continuously tracking whether designated zones remain clear and maintaining accurate enter/exit occupancy counts for spatial analytics and capacity management.

Latency Measurements

Customization

The Blueprint supports several levels of customization:

• Data Level: Add, remove, or replace cameras while maintaining the existing workflow.
• Model Level: Fine-tune the RT-DETR perception model to better suit your use cases.
• Application Level: Build new microservices or applications using the provided APIs and components.
• Microservice Level: Modify existing microservices from source code to extend functionality.

Adding New Cameras or Custom Datasets

Step 1: Prepare video sources

Choose one of the following options to provide video sources for the MV3DT profile:

1. Using recorded videos

   Use this option to run the VSS with local video files instead of live camera feeds. Place your videos in the videos directory under the path configured by VSS_DATA_DIR in $VSS_APPS_DIR/industry-profiles/warehouse-operations/overrides.env.

2. Using live RTSP streams

   Use this option when you want to use live RTSP endpoints. Define each camera in $VSS_APPS_DIR/industry-profiles/warehouse-operations/camera_configs/camera_info.json so the VSS Configurator can read the camera list and generate the required configuration. The following steps apply when using live RTSP streams.

   (i) Create a Sensor Info File (camera_info.json):

   {
     "sensors": [
       {
         "camera_name": "camera-01",
         "rtsp_url": "rtsp://<IP_address:port>/stream1",
         "group_id": "<group_id_1>",
         "region": "<region_1>"
       },
       {
         "camera_name": "camera-02",
         "rtsp_url": "rtsp://<IP_address:port>/stream1",
         "group_id": "<group_id_1>",
         "region": "<region_1>"
       }
     ]
   }

   The camera_name value should be unique for each camera, and rtsp_url should point to the camera stream. The group_id field defines camera groups with overlapping fields of view, which is required for multi-camera 3D tracking and BEV fusion. The region field identifies the camera location or area and is required for MV3DT profile camera configuration.

   Required fields: camera_name, rtsp_url, group_id, region

(ii) Configure VSS Configurator environment variables: Set the following in $VSS_APPS_DIR/industry-profiles/warehouse-operations/overrides.env, which is loaded after warehouse .env and takes precedence:

SENSOR_INFO_SOURCE=file
SENSOR_FILE_PATH=<path_to_camera_info.json>

Step 2: Configure number of streams

Configuring Number of Streams

Number of streams to be processed can be configured in two ways:

1. Static Configuration: Set the NUM_STREAMS environment variable in warehouse overrides.env to specify the desired number of streams for video dataset or RTSP streams.

   NUM_STREAMS=4

   The configured number of streams should be less than or equal to the maximum streams supported by your hardware profile and deployment mode. VSS Configurator can be used to automatically cap the stream count using the formula: final_stream_count = min(NUM_STREAMS, max_streams_supported). When SENSOR_INFO_SOURCE=nvstreamer, the same NUM_STREAMS value is the expected camera count the Sensor Configuration Manager waits for before VST registration. For more details, refer to the How to Count Files Dynamically (Prerequisites) section and NVStreamer Settings in VSS Configurator Documentation.

2. Dynamic Configuration: Use the VSS Configurator’s prerequisite operations to automatically count the number of video files in the recorded videos directory and use that count for configuration updates.

Example: Automatically determine stream count from video files in the dataset directory:

# In blueprint_config.yml
commons:
  # Step 1: Count video files BEFORE variable processing
  prerequisites:
    mv3dt:
      - operation_type: "file_management"
        target_directories:
          - "${VSS_DATA_DIR}/videos/${SAMPLE_VIDEO_DATASET}"
        file_management:
          action: "file_count"
          parameters:
            pattern: "*.mp4"
          output_variable: "available_video_count"  # Stores count (e.g., 6)
  # Step 2: Use the count to compute final stream count
  variables:
    mv3dt:
      # Cap to minimum of: available videos, GPU limit
      - final_stream_count: "min(${available_video_count}, ${max_streams_supported})"
  # Step 3: Use computed variable in config file updates
  file_operations:
    mv3dt:
      - operation_type: "text_config_update"
        target_file: "${DS_CONFIG_DIR}/ds-main-config-mv3dt.txt"
        updates:
          max-batch-size: ${final_stream_count}
          batch-size: ${final_stream_count}

How it works: If your dataset directory has 6 video files and the GPU supports a maximum of 4 streams, the configurator computes: final_stream_count = min(6, 4) = 4. For more details, refer to the How to Count Files Dynamically (Prerequisites) section in VSS Configurator Documentation.

When using a custom recorded-video dataset, NUM_STREAMS depends on the number of video files available in that dataset. Make sure NUM_STREAMS is less than or equal to the number of videos you want to process.

To use a custom dataset:

• For a custom synthetic dataset, use a dataset name that ends in -synthetic (for example, my-warehouse-synthetic). Custom real datasets do not have a suffix requirement.

• Set SAMPLE_VIDEO_DATASET in $VSS_APPS_DIR/industry-profiles/warehouse-operations/overrides.env to the name of your dataset.

• Place the video files in $VSS_DATA_DIR/videos/$SAMPLE_VIDEO_DATASET/.

• Add the image metadata and floor map files under $VSS_APPS_DIR/industry-profiles/warehouse-operations/warehouse-mv3dt-app/calibration/sample-data/${SAMPLE_VIDEO_DATASET}. If you already have a calibration, add calibration.json to the same directory; otherwise, generate it in Step 3. If you already have camInfo files, add them under the camInfo directory; otherwise, create them in Step 4. You can refer to sample datasets provided in the repository for the expected file structure and sample contents.

Step 3: Deploy and generate new calibration

Before deploying the stack and generating a calibration, create the projects directory:

mkdir -p ${VSS_APPS_DIR}/services/auto-calibration/projects

Change the ownership of the Auto Calibration data and projects directories to UID 1000 and GID 1000:

sudo chown -R 1000:1000 ${VSS_DATA_DIR}/auto-calib
sudo chown -R 1000:1000 ${VSS_APPS_DIR}/services/auto-calibration/projects

Configure industry-profiles/warehouse-operations/overrides.env using the Auto-Calibration tab under Deployment Selection, then follow Deploy the Blueprint to start the Auto-Calibration profile. Use it to generate a new calibration against live RTSP streams served by nvstreamer, using the warehouse dataset and VST stack. For steps to generate calibration, see Auto Calibration workflow steps. After generating and storing the calibration, bring the Docker Compose stack down and update the calibration file at $VSS_APPS_DIR/industry-profiles/warehouse-operations/warehouse-mv3dt-app/calibration/sample-data/${SAMPLE_VIDEO_DATASET}/calibration.json.

Step 4: Update MV3DT-specific configurations

After the MV3DT calibration file is generated and placed in the deployment sample-data directory, update the MV3DT-specific configuration files using the utility scripts under tools/rtvi-cv-mv3dt-utils. Complete these updates before running camera clustering.

When adding new cameras to the MV3DT profile, you must update the following MV3DT-specific configuration files:

1. Camera Information Files (camInfo/<sensor_id>.yml): Each camera requires a camInfo file containing the 3x4 projection matrix and per-class object model dimensions. Generate these from your calibration.json using the provided utility script. See Update Camera Information Configuration for details.

2. MQTT Publish/Subscribe Configuration (pub_sub_info_config.yml): Defines the inter-camera communication graph for MV3DT. Generate a vision-neighbor graph from camera calibration data. See 3D Multi Camera Detection and Tracking (MV3DT) for details.

3. Tracker Configuration (ds-mv3dt-tracker-config.yml): Update the ObjectModelProjection.cameraModelFilepath section to map each new sensor ID to its corresponding camInfo file.

Step 5: Camera Clustering

After the MV3DT-specific configuration files are updated, use create_camera_clusters.py to partition cameras into non-overlapping groups for separate 3D model instances. The warehouse MV3DT compose setup uses one BEV group, so run the clustering tool with --n_clusters 1 and verify that sensors[].group is present in calibration.json. For command examples and available options, see Camera Clustering.

Step 6: Redeploy the stack to apply the new calibration

Before redeploying, replace the Auto-Calibration settings in industry-profiles/warehouse-operations/overrides.env with the settings from the MV3DT tab under Deployment Selection. Then follow Deploy the Blueprint to start the MV3DT profile with the new calibration.

docker ps -a

Camera Grouping Utilities

The MV3DT Blueprint uses multi-camera BEV (Bird’s Eye View) perception systems that require cameras to be organized into groups. Each camera group defines a set of cameras with overlapping fields of view that work together for 3D detection and tracking.

The spatialai_data_utils library provides the following functionality:

• Calculate BEV origin and dimensions for each camera group.
• Partition cameras into clusters using spatial clustering algorithms for distributed 3D model deployment.
• Filter and select specific cameras for processing.
• Generate FOV polygons from camera intrinsic/extrinsic matrices.
• Visualize camera coverage and group assignments.

To install the library:

pip install spatialai-data-utils==2.0.1 --extra-index-url=https://edge.urm.nvidia.com/artifactory/api/pypi/sw-metropolis-pypi/simple

The camera grouping tools are shipped with the full spatialai-data-utils package in the VSS repository. The examples below run the scripts from the package’s tools/camera_grouping directory.

BEV Group Origin Calculation

Use calculate_origin.py to calculate BEV origin and dimensions for camera groups, filter cameras, and generate FOV visualizations.

The algorithm works as follows:

1. FOV Calculation: By default, the tool uses frustum-based FOV generation. It projects camera view frustums onto the ground plane using camera intrinsic (focal length, principal point) and extrinsic (rotation, translation) matrices. The frustum is intersected with a configurable height range to create a polygon representing the camera’s ground coverage.
2. Group Bounds: For each camera group, the tool computes the union of all camera FOV polygons and calculates the bounding box with optional dilation.
3. Origin Calculation: The BEV origin is computed as the centroid of the union FOV polygon, and dimensions are derived from the bounding box of the union, providing the coordinate reference for multi-camera tracking.

Origin calculation is required when setting up a new multi-camera deployment or when camera configurations change. This functionality is already integrated into the VSS Configurator, so manual execution is typically not necessary. However, if you need to run it as a standalone script for debugging, custom workflows, or batch processing, the example commands below demonstrate the available options.

Navigate to the spatialai-data-utils camera grouping tools

export VSS_REPO_DIR=/path/to/video-search-and-summarization
cd $VSS_REPO_DIR/libs/analytics/spatialai-data-utils/tools/camera_grouping/

# Using dataset folder (auto-detects calibration.json)
python calculate_origin.py data/scene
# Using direct path to calibration file
python calculate_origin.py data/scene/calibration.json
# Custom height range for ground plane intersection
python calculate_origin.py data/scene/calibration.json --height-range 0.5 2.5
# Process only specific sensors
python calculate_origin.py data/scene/calibration.json --sensor-names Camera1,Camera2,Camera3
# Specify output file
python calculate_origin.py data/scene/calibration.json -o calibration_with_origins.json
# Overwrite the original calibration file
python calculate_origin.py data/scene/calibration.json --overwrite
# Include visualization (uses black background if map file not provided)
python calculate_origin.py data/scene/calibration.json --visualize
# Include visualization with map file
python calculate_origin.py data/scene/calibration.json --map_file data/scene/Top.png --visualize
# Use existing FOV polygons from calibration instead of frustum calculation
python calculate_origin.py data/scene/calibration.json --prefer-existing-fov
# Constrain FOV with scene bounds
python calculate_origin.py data/scene/calibration.json \
    --scene-bounds -30 -40 30 40 --max-camera-distance 25.0

Use --help to list all available arguments.

Key options:

• --output, -o: Output calibration file path (default: input_with_origins.json).
• --overwrite: Overwrite the input calibration file (mutually exclusive with --output).
• --map_file: Path to map image for visualization (uses black background if not provided).
• --sensor-names: Filter to process only specified sensor names (comma-separated).
• --n-sensor-groups: Number of sensor groups to create when group field is missing (default: 1).
• --dilation: Dilation distance in meters for group bounds calculation (default: 1.0).
• --height-range: Height range (min, max) in meters for ground plane intersection (default: 1.0 3.0).
• --max-camera-distance: Maximum distance in meters to constrain frustum polygons (default: 30.0).
• --scene-bounds: Scene bounds (min_x, min_y, max_x, max_y) in meters to clip frustum polygons.
• --prefer-existing-fov: Use existing FOV from calibration file, fall back to frustum if not available.
• --visualize: Generate visualization of groups.
• --vis_separate_images: Generate separate visualization images per group instead of combined.

Camera Clustering

For 3D model inference and deployment, cameras need to be partitioned into non-overlapping groups where each group is handled by a separate 3D model instance. Use create_camera_clusters.py to partition cameras into spatially compact clusters based on FOV coverage and spatial proximity. This tool assigns ALL cameras to exactly N clusters, ensuring efficient distribution of camera workloads across multiple model instances.

Running Camera Clustering

Navigate to the spatialai-data-utils camera grouping tools

export VSS_REPO_DIR=/path/to/video-search-and-summarization
cd $VSS_REPO_DIR/libs/analytics/spatialai-data-utils/tools/camera_grouping/

# Basic usage with densify mode (default, visualization on by default)
python create_camera_clusters.py data/scene --max_camera_per_group 10
# Use balanced mode for evenly distributed clusters
python create_camera_clusters.py data/scene --max_camera_per_group 10 --mode balanced
# Densify mode with custom thresholds (disabling auto-tuning)
python create_camera_clusters.py data/scene --max_camera_per_group 10 --mode densify \
    --min_overlap_threshold 0.3 --max_distance_threshold 10.0 --disable_param_tuning
# Balanced mode with specific number of clusters
python create_camera_clusters.py data/scene --max_camera_per_group 8 --mode balanced \
    --n_clusters 5
# Specify output path and overwrite input file
python create_camera_clusters.py data/scene --max_camera_per_group 10 \
    --output data/scene/calibration_clustered.json --overwrite
# Use existing FOV polygons from calibration instead of frustum calculation
python create_camera_clusters.py data/scene --max_camera_per_group 10 --prefer_existing_fov
# Custom auto-tuning search grids
python create_camera_clusters.py data/scene --max_camera_per_group 10 \
    --tuning_overlap_grid 0.1 0.2 0.3 --tuning_distance_grid 5.0 8.0 10.0

Use --help to list all available arguments.

Key options:

• --max_camera_per_group: Maximum cameras per cluster (required). Has higher priority than --n_clusters.
• --n_clusters: Override auto-calculated number of clusters.
• --mode: Clustering mode - “balanced” or “densify” (default: densify).
• --output: Output path for the clustered calibration file.
• --overwrite: Overwrite the input calibration file.
• --output_suffix: Suffix for output files (default: “clustered”).
• --start_camera_index: Starting camera index for seeding (default: 0).
• --min_overlap_threshold: Minimum required FOV overlap (default: 0.2).
• --max_distance_threshold: Maximum allowed centroid distance in meters (default: 8.0).
• --max_cascade_depth: Maximum recursion depth for densify-mode cascade reassignment (default: 3).
• --prefer_existing_fov: Use existing FOV polygons in calibration instead of calculating from frustum.
• --height_range: Height range (min, max) in meters for ground plane intersection (default: 1.0 3.0).
• --image_size: Image dimensions (width, height) in pixels for frustum calculation (default: 1920 1080).
• --max_camera_distance: Maximum distance in meters for frustum calculation (default: 30.0).
• --dilation: Buffer distance in meters for cluster bounding boxes (default: 8.0).
• --disable_param_tuning: Disable auto-tuning of clustering parameters.
• --tuning_overlap_grid: Custom overlap thresholds (0-1) to search when auto-tuning.
• --tuning_distance_grid: Custom centroid distance thresholds (meters) to search when auto-tuning.
• --tuning_start_index_grid: Seed camera indices to try when auto-tuning.
• --tuning_workers: Number of parallel workers for auto-tuning (0=auto, 1=disable parallelism).
• --vis_no_camera_id_labels: Disable drawing camera IDs on the visualization.
• --vis_separate_images: Generate separate visualization images per cluster instead of combined.

Clustering Modes

• densify (default): Prioritizes creating full, densely-packed clusters. Uses cascade reassignment to handle unassigned cameras by recursively attempting to place them in nearby clusters.
• balanced: Enforces strict thresholds for overlap and distance. Splits oversized clusters to maintain balance across all groups.

Cluster Count Calculation

The number of clusters is determined by the following logic:

1. If --n_clusters is not provided, it is auto-calculated as ceil(total_cameras / max_camera_per_group).
2. If --n_clusters is explicitly set but would violate the --max_camera_per_group constraint (i.e., requires more clusters than specified), the tool overrides --n_clusters with the minimum required value and logs a warning.
3. If --n_clusters=1 is set but total_cameras > max_camera_per_group, an error is raised.

This ensures --max_camera_per_group always takes priority to guarantee each cluster stays within the specified camera limit.

Finding Optimal Clustering Parameters

Use find_suggested_cluster_params.py to search for optimal clustering parameters. This tool performs a grid search over overlap thresholds, distance thresholds, and seed camera indices to find settings that produce compact, capacity-respecting clusters:

# Basic usage - find best parameters for clustering
python find_suggested_cluster_params.py data/scene --max_camera_per_group 10
# Custom search grids
python find_suggested_cluster_params.py data/scene --max_camera_per_group 10 \
    --overlap_grid 0.1 0.2 0.3 --distance_grid 5.0 8.0 10.0
# Specify seed camera indices to try
python find_suggested_cluster_params.py data/scene --max_camera_per_group 10 \
    --start_index_grid 0 5 10
# Use random sampling for start indices
python find_suggested_cluster_params.py data/scene --max_camera_per_group 10 \
    --start_index_seed 42
# Show more candidates with verbose output
python find_suggested_cluster_params.py data/scene --max_camera_per_group 10 \
    --top_k 10 --verbose

Key options:

• --max_camera_per_group: Maximum cameras per cluster (required).
• --mode: Clustering mode to evaluate - “balanced” or “densify” (default: densify).
• --prefer_existing_fov: Use existing FOV polygons instead of calculating from frustum.
• --max_camera_distance: Maximum distance in meters for frustum calculation (default: 30.0).
• --height_range: Height range (min, max) in meters for ground plane intersection (default: 1.0 3.0).
• --image_size: Image dimensions (width, height) in pixels for frustum calculation (default: 1920 1080).
• --max_cascade_depth: Maximum recursion depth for densify-mode cascade (default: 3).
• --overlap_grid: List of overlap thresholds (0-1) to search.
• --distance_grid: List of centroid distance thresholds (meters) to search.
• --start_index_grid: Seed camera indices to try.
• --start_index_seed: Random seed for auto-generated start camera indices.
• --top_k: Number of top candidates to display (default: 5).
• --workers: Number of parallel workers for the sweep (0=auto, 1=disable parallelism).
• --verbose: Enable verbose logging.

The output ranks candidates by score (lower is better), showing unassigned camera count, overflow, and scatter metrics for each parameter combination.

Camera Group Reassigning

Use reassign_camera_groups.py to manually reassign cameras to different existing BEV groups. Common use cases include:

• Fixing suboptimal clustering: Move cameras that were incorrectly assigned by the automatic clustering algorithm.
• Balancing cluster sizes: Redistribute cameras between groups to achieve more even workload distribution.
• Iterative refinement: Fine-tune cluster assignments after reviewing visualization results.

Navigate to the spatialai-data-utils camera grouping tools

export VSS_REPO_DIR=/path/to/video-search-and-summarization
cd $VSS_REPO_DIR/libs/analytics/spatialai-data-utils/tools/camera_grouping/

# Move cameras to different groups
python reassign_camera_groups.py data/calibration.json \
    --move cam-01:bev-sensor-2 cam-05:bev-sensor-3
# Specify output path
python reassign_camera_groups.py data/calibration.json \
    --move cam-01:bev-sensor-2 --output data/calibration_updated.json
# Overwrite the original file
python reassign_camera_groups.py data/calibration.json \
    --move cam-01:bev-sensor-2 --overwrite
# Strict mode - fail if camera or group is missing
python reassign_camera_groups.py data/calibration.json \
    --move cam-01:bev-sensor-2 --strict
# Use existing FOV polygons instead of frustum calculation
python reassign_camera_groups.py data/calibration.json \
    --move cam-01:bev-sensor-2 --prefer_existing_fov

Key options:

• --move: Mappings of camera_id:group_name (space separated) to reassign (required).
• --output: Output path for updated calibration (default: <input>_reassigned.json).
• --overwrite: Overwrite the input calibration file in-place.
• --output_suffix: Suffix for output files (default: “reassigned”).
• --strict: Fail if a camera or target group is missing; otherwise skip with warning.
• --prefer_existing_fov: Use existing FOV from calibration, fall back to frustum if not available.
• --map_file: Path to map image for visualization (auto-detects Top.png if omitted).
• --dilation: Dilation distance in meters when recomputing group bounds (default: 1.0).
• --height_range: Height range (min, max) in meters for ground plane intersection (default: 1.0 3.0).
• --image_size: Image dimensions (width, height) in pixels for frustum calculation (default: 1920 1080).
• --max_camera_distance: Maximum distance in meters for frustum calculation (default: 30.0).
• --vis_no_camera_id_labels: Disable drawing camera IDs on the visualization.

Model Customization

Perception Model Fine-tuning

The MV3DT profile uses RT-DETR (Real-Time Detection Transformer) as its 2D perception model and the MV3DT module for 3D multi-camera tracking. To customize:

1. Fine-tune RT-DETR using TAO (RT-DETR model page)
2. Configure DeepStream to use your custom model (model configuration guide)

For detailed instructions on integrating a new RT-DETR model with the MV3DT pipeline, including model compatibility requirements and configuration steps, refer to the MV3DT Perception Microservice documentation.

Application Customization

The Blueprint uses a modular microservices architecture with the following communication channels:

• Message Broker (Kafka or Redis)
• Elasticsearch database
• REST APIs

Users can build their own microservices by consuming data from the above channels.

For complete API documentation, see API Reference Page.

Available Service Ports

The following ports are used during deployment, and users can leverage them for any potential integration:

Analytics Microservices Customization (Advanced)

For detailed information about customizing specific analytics microservices, refer to:

• Behavior Analytics
• VSS Video Analytics API

For detailed information about customizing the MV3DT perception pipeline, including DeepStream configuration, tracker parameters, camera information files, and MQTT publish/subscribe topology, refer to the MV3DT Perception Microservice documentation.

Hardware Config Customization

The MV3DT Warehouse Blueprint requires several configuration files to be properly tuned based on your GPU hardware and deployment requirements. When changing hardware (e.g., switching from H100 to L4 GPU) or adjusting the number of video streams, multiple configuration files must be updated to ensure optimal performance and prevent GPU overload.

Configuration Files Requiring Hardware-Based Updates

The following table lists the configuration files that typically require updates when hardware changes:

Certain GPUs require additional optimizations. For example, IGX-THOR and DGX-SPARK get:

• Extended batched-push-timeout (67000ms) for DeepStream configs
• enable_overlay_skip_frame enabled in VST configs for better performance

There are two approaches to customize these configuration files:

Approach 1: Manual Configuration

Manually update all required configuration files before deploying the blueprint. This is time consuming, error prone and often not suitable for production deployments.

# Update DeepStream main config
vi <PATH_TO_DS_CONFIG_DIR>/ds-main-config-mv3dt.txt
# Set: max-batch-size=<stream_count>, batch-size=<stream_count>
# Update NvStreamer config
vi <PATH_TO_NVSTREAMER_CONFIG_DIR>/vst-config.json
# Set: "nv_streamer_sync_file_count": <stream_count>, "max_devices_supported": <max_streams>
# Update VIOS (VST) config
vi <PATH_TO_VST_CONFIG_DIR>/vst_config.json
# Set: "onvif.max_devices_supported": <max_streams>, "data.always_recording": true,
#      "notifications.use_message_broker_consumer": "<deployment_mode>"

Approach 2: Automatic Config Management using VSS Configurator

The VSS Configurator provides a declarative approach to automatically update all required configuration files based on your hardware profile and deployment mode. This is the recommended approach for production deployments and simplifies the configuration management process.

For how the configurator updates configurations in this blueprint, see VSS Configurator for the Warehouse Blueprint. The guide describes configuration inputs, affected files, stream ceilings, and hardware-specific overrides.

To enable the VSS Configurator, in VSS Configurator’s environment variables you must set:

ENABLE_PROFILE_CONFIGURATOR=true

The VSS Configurator is enabled by default (ENABLE_PROFILE_CONFIGURATOR=true). When enabled, it runs before the Blueprint Deployment starts and adjusts configuration files based on the hardware profile and deployment mode as defined in the HARDWARE_PROFILE and MODE environment variables.

The VSS Configurator provides a comprehensive set of features for automated profile configuration management:

Execution Order: Environment Variable Validation → Prerequisite Operations → Variable Computations → Configuration File Updates

For detailed information on how to create custom hardware profiles and advanced configuration options, refer to the Profile Configuration Manager section in VSS Configurator Documentation.