Please provide complete information as applicable to your setup.
• Hardware Platform (Jetson / GPU) 3090RTX
• DeepStream Version 9.0
• JetPack Version (valid for Jetson only)
• TensorRT Version deepstream 9.0 installation guide
• NVIDIA GPU Driver Version (valid for GPU only) 590
• Issue Type( questions, new requirements, bugs)
• How to reproduce the issue ? (This is for bugs. Including which sample app is using, the configuration files content, the command line used and other details for reproducing)
• Requirement details( This is for new requirement. Including the module name-for which plugin or for which sample application, the function description)
Hello,
I am using NVIDIA DeepStream with the nvinfer plugin for object detection on MP4 video.
My model was trained for aerial imagery and expects relatively small inputs, for example 300×300 pixels. For large video frames such as 1920×1080, I would like to divide each frame into multiple overlapping tiles, run inference on every tile, convert each detection back to the original frame coordinates, and then apply NMS to remove duplicate detections from overlapping tiles.
The intended workflow is:
-
Split the original frame into overlapping 300×300 tiles.
-
Run the same detector on each tile.
-
Convert bounding boxes from tile coordinates to full-frame coordinates.
-
Merge all detections.
-
Apply class-wise NMS to remove duplicates.
Does the standard DeepStream nvinfer plugin support this type of tiled inference directly?
For example, is there any configuration option in nvinfer for:
-
splitting one frame into several overlapping crops,
-
batching those crops for inference,
-
restoring detections to the original frame coordinates,
-
and merging duplicate detections?
Or do I need to implement this manually using a custom GStreamer plugin, nvdspreprocess, custom metadata, or multiple nvinfer inputs?
I would also appreciate any recommended DeepStream architecture for implementing tiled inference efficiently on GPU.
Thank you.