deepstream-import-vision-model

To convert a HuggingFace object detection model to a DeepStream pipeline, you need to export SafeTensors to ONNX, build a TensorRT engine, and compile a custom nvinfer bbox parser for inference.

Benchmarking a TensorRT engine involves running trtexec with fixed log filenames to measure throughput and derive capacity. This validates engine performance prior to integration into the DeepStream pipeline.

Yes, you can use custom nvinfer bbox parsers with dynamically named TensorRT engines. The parser must be generated and compiled specifically for the dynamically built engine to ensure correct detection output parsing.

Validating ONNX model detection results against the KITTI dataset requires running KITTI-based detection validation. This process compares parsed bounding box outputs against ground truth labels to verify model accuracy.

Generating an HTML and PDF benchmark report for DeepStream performance requires running single-stream and multi-stream measurements. The md-to-html-pdf.py script then converts these fixed logs into a styled canonical report.

Yes, DeepStream deployment requires extracting class labels from the model's config.json. This ensures the custom nvinfer bbox parser correctly maps TensorRT engine detection outputs to the intended object categories.