> This page is for version 3.3.0 (default). > For other versions, use one of these documentation indexes: > - 3.3.0 (default): https://docs.nvidia.com/vss/3.3.0/llms.txt > - Dev: https://docs.nvidia.com/vss/dev/llms.txt > For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.nvidia.com/vss/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/vss/_mcp/server. # Known Limitations This section aligns with PRD and current release constraints, providing transparency about known limitations. ## Perception ### Truncated Bounding Box In the 2D Blueprint, if an object is partially visible due to occlusion or being outside the camera's field of view, the detected bounding box from the perception model can be truncated. The midpoint on the lower edge of the bounding box is used to approximate the object's location, which is then used for various downstream analytics computations such as proximity detection. If the bounding box is truncated, the accuracy of location-related analytics can be impacted. For example, in the image below, the person on the right has a truncated bounding box because they are partially outside the camera's field of view. This can affect the accuracy of proximity detection between the two individuals. ![Truncated Bbox Proximity Detection](/vss/_fern-img/d25fac38dc6fb4347a48a63ecfc004dafba4f6c292f2a49800a1a6eada3982ad.webp) More advanced techniques will be included in future releases to address similar issues. ### False Object Detections in MV3DT The [MV3DT Vision AI Profile](/vision-agent/agent-workflows/industry-specific-examples/warehouse-operations-blueprint/development-workflow/deployment-customization/mv3dt-vision-ai-profile) may exhibit the following detection and tracking issues: * **Tracker ID switches:** When people move close together, a tracker ID may switch from one person to another. IDs may also become associated with different objects across camera views. * **Tracking loss:** Previously tracked objects may lose their tracks while still visible. * **False detections:** Bounding boxes may appear on background structures or empty areas where no target object is present. * **Detached or oversized bounding boxes:** Boxes may be offset from their objects, float above them, or extend substantially beyond their dimensions. * **Duplicate bounding boxes:** Multiple boxes, sometimes with different tracker IDs, may appear on the same object. * **Unstable bounding boxes:** Boxes may flicker or change inconsistently between frames. * **Missed detections:** Visible people or vehicles may have no bounding box. ### Appearance ReID gallery and GPU * The Milvus gallery is not durable storage. Warehouse Compose has no named volumes for etcd, MinIO, or Milvus, so `docker compose restart` keeps appearance memory, and `docker compose down` (or recreating those containers) empties it until new `mdx-raw` embeddings are ingested. Standalone named volumes survive `docker compose down` unless you pass `-v`. Helm uses `emptyDir` for those backends, so the gallery is cleared when the pods are recreated, not on a process restart inside a running pod. * Re-association applies to **Person** only (`operateOnClassIds: [0]`). * Helm files three GPU claims (perception, stream processing, `reid-embed`) against two physical cards. Without GPU time-slicing, first install can deadlock: the init Job needs a GPU to export CLIP-ReID while other pods already hold both cards. Enable time-slicing or stage models out of band as described in the [warehouse-mv3dt-app Helm README](https://github.com/NVIDIA-AI-Blueprints/video-search-and-summarization/blob/v3.3.0/deploy/helm/industry-profiles/warehouse-operations/warehouse-mv3dt-app/README.md). ## Helm Deployment ### VST Software Mode Without a GPU Can Affect Overlays and the Video Wall Use a dedicated GPU for VIOS streamprocessing in the Warehouse 2D, 3D, and MV3DT Helm charts. Do not use software mode: it can cause video bounding-box overlay (VBO) misalignment and video-wall failures. Keep hardware video processing enabled and allocate one GPU to streamprocessing: ```yaml vios: vss-vios-streamprocessing: useSoftwarePath: false resources: limits: nvidia.com/gpu: 1 requests: nvidia.com/gpu: 1 ``` RTVI-CV retains its own GPU allocation. ### GPU Sharing If there are not enough physical GPUs for dedicated allocations, configure GPU sharing before deployment while keeping hardware video processing enabled. [Multi-Instance GPU (MIG)](https://docs.nvidia.com/datacenter/cloud-native/gpu-operator/latest/gpu-operator-mig.html) provides memory and fault isolation on supported hardware; [GPU time-slicing](https://docs.nvidia.com/datacenter/cloud-native/gpu-operator/latest/gpu-sharing.html) shares GPU access without that isolation. Choose based on your hardware and workload requirements, including VIOS video encode/decode support in the selected MIG profile. See the [MIG and time-slicing comparison](https://docs.nvidia.com/datacenter/cloud-native/gpu-operator/latest/gpu-sharing.html#comparison-time-slicing-and-multi-instance-gpu). GPU time-slicing can be used with the Warehouse Blueprint’s [2D](/vision-agent/agent-workflows/industry-specific-examples/warehouse-operations-blueprint/development-workflow/deployment-customization/2d-vision-ai-profile), [3D](/vision-agent/agent-workflows/industry-specific-examples/warehouse-operations-blueprint/development-workflow/deployment-customization/3d-vision-ai-profile), [MV3DT](/vision-agent/agent-workflows/industry-specific-examples/warehouse-operations-blueprint/development-workflow/deployment-customization/mv3dt-vision-ai-profile) and [2D with Agents](/vision-agent/agent-workflows/industry-specific-examples/warehouse-operations-blueprint/development-workflow/deployment-customization/2d-vision-ai-with-agents-profile) profiles. For ConfigMap and device-plugin configuration examples, see [NVIDIA GPU time-slicing configuration](https://docs.nvidia.com/datacenter/cloud-native/gpu-operator/latest/gpu-sharing.html#configuration). Two shared allocations allow RTVI-CV and VIOS alone to share one physical GPU. Additional GPU workloads, such as ReID or in-cluster RT-VLM, require additional shared allocations and sufficient GPU memory. ## VLM Verification of Incidents When a person is inside a forklift, the interaction may be incorrectly classified as a near-miss alert. ## Agents See [Agent Known Issues](/vision-agent/vss-agents/agent-known-issues) for agent limitations. ## Alert Verification ### RTX PRO 4500: Dropped Live VLM Results With `MODE=2d` and `BP_PROFILE=bp_wh`, RTVI-VLM running on a single RTX PRO 4500 Blackwell GPU may drop video chunks and result messages when the live workload saturates that GPU. This can cause missing VLM responses and alerts even while the container reports healthy. Logs may show `Live decoder backlog exceeded` or `Kafka async send queue is full`. To reduce load, remove or comment out unneeded rule blocks under `always_on_rules` in `$VSS_APPS_DIR/industry-profiles/warehouse-operations/vlm-as-verifier/realtime-config.yml`, keeping at least one rule, then redeploy the stack. Removed rules will no longer generate alerts. ### VLM Connectivity Errors While running the Warehouse Blueprint **2D Vision AI with Agents Profile**, you may sometimes see the following error in mdx-vlm-incidents: `VLM analysis error: cannot access local variable 'response_content' where it is not associated with a value` This can occur when there is a connectivity issue with the VLM service. ## Camera Calibration with Auto Calibration The following are known limitations when using the [Auto Calibration service](/system-components/calibration/auto-calibration). ### Floor Map and Scale Factor Requirement Auto Calibration currently requires the user to provide a floor map and a proper `scaleFactor`. Auto Calibration uses these inputs to scale the global coordinate system to meters. Behavior analytics assumes the global coordinate system is in meters and computes distance-related properties, such as proximity violation, based on it. If the `scaleFactor` is not provided or is improper with respect to the floor map, the computed distance-related properties can be affected. In future releases, Auto Calibration will remove the requirement to provide both a floor map and `scaleFactor`. ### AMC requires detected person movement The tracklet-based AMC method calibrates cameras from detected person movements in input videos. It requires sufficient people moving through different areas of each camera's view; footage without enough person movement is unsuitable for AMC. This requirement does not apply to VGGT. For multi-camera calibration with VGGT support installed, users can choose whether to run AMC or VGGT first; VGGT does not require a successful AMC run. ### Tripwires Limited to Two Points The Auto Calibration UI supports only a two-point tripwire: a start point and an end point. Users cannot add intermediate vertices to create a polyline. UI annotations represent a single line segment, and the raw exported calibration stores only `wire.p1` and `wire.p2`. Polyline tripwire authoring and export are currently unsupported. ## Brev Launchable The following limitations apply when running the Warehouse Blueprint via the [Brev Launchable](/vision-agent/agent-workflows/industry-specific-examples/warehouse-operations-blueprint/quickstart#brev-launchable): * **Local LLM** — Local LLM deployment has not been tested or verified on Brev. Use a remote LLM when running the 2D with Agent profile.