nvmmconvert

VIC-accelerated 2D transform on NVMM surfaces: crop, scale, color-format convert, rotate/flip, all GPU-to-GPU (no CPU on the data path). Runs on the Tegra VIC via NvBufSurfTransform.

Sink/src caps: video/x-raw(memory:NVMM), formats {NV12, RGBA, I420, BGRA}, up to 8192×8192.

Properties

Property	Type	Default	Description
crop-x	uint	0	Source crop X offset (pixels)
crop-y	uint	0	Source crop Y offset
crop-w	uint	0	Source crop width (0 = full)
crop-h	uint	0	Source crop height (0 = full)
flip-method	enum	0	0=none, 1=rotate-90 (CW), 2=rotate-180, 3=rotate-270 (CCW), 4=horizontal-flip, 5=transpose, 6=vertical-flip, 7=inverse-transpose
interpolation	enum	6	VIC scaling filter: 0=nearest, 1=bilinear, 2=5-tap, 3=10-tap, 4=smart, 5=nicest, 6=default
compute-mode	enum	0	Engine that runs the transform: 0=default (VIC on Tegra), 1=gpu, 2=vic

Output dimensions come from the downstream caps. rotate-90/270 and transpose/inverse-transpose swap width and height.

Examples

# Scale 1080p -> 480p with the 5-tap filter
... ! nvmmconvert interpolation=5-tap ! 'video/x-raw(memory:NVMM),width=640,height=480' ! ...

# Force the GPU engine instead of the VIC (e.g. when the VIC is saturated)
... ! nvmmconvert compute-mode=gpu ! 'video/x-raw(memory:NVMM),width=640,height=480' ! ...

# Crop a region
... ! nvmmconvert crop-x=100 crop-y=50 crop-w=800 crop-h=600 ! \
      'video/x-raw(memory:NVMM),width=800,height=600' ! ...

# Rotate 90 CW (640x480 -> 480x640)
... ! nvmmconvert flip-method=rotate-90 ! 'video/x-raw(memory:NVMM),width=480,height=640' ! ...

nvmmconvert covers the same 2D operations as NVIDIA's stock nvvidconv, as an open-source element integrated with this suite's NVMM allocator and IPC pool.

DeepStream interop

nvmmconvert emits video/x-raw(memory:NVMM) backed by NvBufSurface (NV12 / RGBA), the same memory DeepStream consumes, so it works as a preprocessing stage (crop / scale / rotate on the VIC) ahead of inference. Place it before nvstreammux — nvinfer runs on the batched mux output, so per-stream ROI crops or resizes belong upstream of the mux:

Verified on Orin

Tested on Orin NX (JP6) inside the nvcr.io/nvidia/deepstream:7.1-samples container: the pipeline below ran end-to-end with real inference (TrafficCamNet). nvstreammux property names and the nvinfer config layout vary by DeepStream release — adjust config-file-path to your install.

# Crop a region of interest + scale, then batch and infer
gst-launch-1.0 \
  filesrc location=video.mp4 ! qtdemux ! h264parse ! nvv4l2decoder \
  ! 'video/x-raw(memory:NVMM)' \
  ! nvmmconvert crop-x=320 crop-y=180 crop-w=1280 crop-h=720 \
  ! 'video/x-raw(memory:NVMM),width=640,height=640,format=NV12' \
  ! nvstreammux batch-size=1 width=640 height=640 \
  ! nvinfer config-file-path=pgie.txt batch-size=1 \
  ! nvvideoconvert ! nvdsosd ! nv3dsink

Set nvinfer batch-size=1 for a single stream

The stock DeepStream config_infer_primary.txt defaults to a batch-30 INT8 engine (built for the multi-stream deepstream-app). Building that engine needs far more GPU memory than one stream warrants and can OOM on an 8 GB Orin NX. Overriding nvinfer batch-size=1 builds a small single-batch engine instead.

nvmmconvert is interchangeable with nvvideoconvert for this 2D step; it additionally shares this suite's NVMM allocator and IPC pool, so a cropped frame can be published over IPC without leaving the GPU.