skills.rest
NVIDIA
NVIDIAOFFICIAL
Explore all skills by this author →
WebsiteGitHub

rtvi-cv-customize-model

Replace the DeepStream CV detection model in VSS Alerts Blueprint verification mode with a custom ONNX detector.

3.2k|370|Updated Feb 25, 2026
Search Tags:#docker compose#object detection#ONNX#tensorRT#vss#DeepStream#nvinfer
One-click install
npx skills add https://github.com/NVIDIA/skills --skill rtvi-cv-customize-model
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: rtvi-cv-customize-model
Source: https://github.com/NVIDIA/skills/tree/main/skills/rtvi-cv-customize-model
Command: npx skills add https://github.com/NVIDIA/skills --skill rtvi-cv-customize-model

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

Swapping the stock DeepStream detector in the NVIDIA VSS Alerts Blueprint verification (2d_cv) mode involves many fragile steps — ONNX staging, custom bbox parser compilation, nvinfer configuration, runtime TensorRT engine builds, and compose mount pitfalls — where small mistakes cause silent failures like ghost directories or dlsym parser errors.

Core Features & Use Cases

  • End-to-End Model Swap Workflow: Guides ONNX staging under VSS_DATA_DIR, custom bbox parser authoring and compilation, nvinfer config (cluster-mode, parse-bbox-func-name), runtime TRT engine build via ds-start.sh, and redeployment of the perception-alerts service.
  • Troubleshooting Guidance: Covers common failures such as Docker ghost-directory mounts, dlsym parser symbol mismatches, dev-profile.sh wiping staged models, and single-GPU device reservation conflicts.
  • Segmentation Model Addendum: Defines the frame-mask contract for detection-plus-mask or instance-segmentation models before parser or handoff code is written.
  • Use Case: Replace the default GDINO/RTDETR detector with a YOLOv11 COCO ONNX model, build the TensorRT engine at container startup, and redeploy perception-alerts on a single-GPU host.

Quick Start

Ask your agent to replace the VSS Alerts Blueprint verification-mode detector with your custom ONNX model and redeploy the perception-alerts service.

Frequently Asked Questions about rtvi-cv-customize-model

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I replace the VSS Alerts Blueprint detection model with a custom ONNX model?▼

Stage your ONNX model under ${VSS_DATA_DIR}/models/yolo, create a custom bbox parser and nvinfer config, add a runtime TensorRT engine build block to ds-start.sh, then update compose and redeploy the perception-alerts service. The skill walks through each step using YOLOv11 COCO 80 as the worked example.

Why does DeepStream report dlsym failed for my custom bbox parser?▼

The parse-bbox-func-name in the nvinfer config, the extern "C" function symbol in your parser source, and the CHECK_CUSTOM_PARSE_FUNC_PROTOTYPE macro must all match exactly. Also rebuild the .so with both DeepStream and CUDA include paths present.

Does this skill work with VSS real-time alerts (2d_vlm) mode?▼

No. This skill only applies to verification mode (2d_cv), which runs a DeepStream CV detector. Real-time alerts mode (2d_vlm) has no CV detector to swap, and scaffolding a new RTVI CV microservice uses a different skill.

Can I mount the ONNX file path directly in the Docker compose file?▼

No. If the host file does not exist yet, Docker creates a ghost directory at that path. Mount the parent directory instead, keep the ONNX under ${VSS_DATA_DIR}/models/yolo, and re-stage it after any dev-profile.sh up run that recreates the models directory.

How do I handle instance segmentation or detection-plus-mask models?▼

Complete the segmentation model contract reference before writing parser or handoff code. It requires declaring the frame-level mask encoding, coordinate space, overlap policy, and handoff path, since a bbox parser alone does not carry masks downstream.

Why does my customized ds-start.sh never run after redeployment?▼

Stock Alerts compose runs services/rtvi/rtvi-cv/ds-start.sh via an extends bind mount, not the profile init-scripts copy. You must override that bind mount in the profile compose file so your edited script is the one mounted into the container.

Related Skills

Explore More Software Engineering Skills→
nishide-devnishide-dev

ml-model-export

Export trained PyTorch models to ONNX, TorchScript, and TensorRT formats.

Community
Advanced
NVIDIANVIDIA
189

amc-run-rtsp-calibration

Automate camera calibration from live RTSP streams via the AutoMagicCalib REST API.

Official
Advanced
NVIDIA-AI-IOTNVIDIA-AI-IOT
81

deepstream-import-vision-model

Automates DeepMind model deployment from neural network definition to hardware-efficient binary.

Official
Advanced
NVIDIANVIDIA
49

vss-sop-build

Build and validate a customized VSS SOP monitoring blueprint from VSS 3.1.

Official
Advanced