INT8 Precision for a CUSTOM YOLOLIKE BOX OBJECT Model

Software
,
1

Hi there,

I currently use the ZED SDK 5.0 in Python with both depth sensing and object detection enabled. My object detection uses CUSTOM_YOLOLIKE_BOX_OBJECTS and a custom onnx file, which I export using Ultralytics’ export function as follows:

from ultralytics import YOLO
YOLO("..path/to/my/model.pt").export(
  format="onnx",
  simplify=True,
  dynamic=False,
  opset=13,
  imgsz=608
)

I would like to improve my ZED performance. Currently, we are using fp16. How can I use int8 precision with my custom onnx model? I see in the ZED documentation that int8 is supported using allow_reduced_precision_inference = True on my sl.ObjectDetectionParameters() object. I am using this but I’m still in FP16.

How can I use int8 precision with a custom yololike box object model? Do I need to export my onnx in a specific way?

Thanks!

2

Hi @stripedapus
I’m checking this problem with the ZED SDK team, and I will be back to you as soon as possible.

Meanwhile, can you provide more information regarding your setup?
What GPU model are you using?

3

Hi, no it’s not supported. The allow_reduced_precision_inference param is currently implemented and active only when using body tracking models.