[Feature Request] NVMM-Compliant CUDA ROI Blur Plugin for DeepStream

Hardware Platform: dGPU
DeepStream Version: 7.x
TensorRT Version: 10.x
NVIDIA GPU Type: dGPU
Operating System: Ubuntu 22.04 (DeepStream Docker container)
Issue Type: Feature Request


Hi DeepStream team,

I’d like to request native support for ROI-level Gaussian blur in DeepStream.

Currently, nvdsosd only supports solid-color rectangle fills for redaction. For de-identification use cases (faces, license plates, GDPR compliance), blurring is strongly preferred over black boxes. The common workaround using pyds.get_nvds_buf_surface() + OpenCV breaks the NVMM zero-copy pipeline by copying pixel data to CPU, which significantly degrades performance.

I built a CUDA-based ROI blur library that operates directly on NvBufSurface GPU memory (NVMM-compliant, no CPU round-trip) as a workaround. Here is an example implementation I developed with GitHub Copilot:

It would be great if the DeepStream SDK could provide a built-in blur option natively, for example as a blur draw mode in nvdsosd or as a dedicated element like nvdsblur, so that the community doesn’t have to build custom CUDA plugins for such a common use case.

Thank you.

Best regards

The objects blur function has been added to DeepStream SDK 8.0 GA. Please check the nvdsosd plugin. Gst-nvdsosd — DeepStream documentation

1 Like

Good news!! Thank you very much.