Not sure why preprocessing time becomes longer in some cases

[IASUI]

When I perform inference on a common 1080p video, the length of pre-processing, inference, and post-processing time is consistent with the performance of using TensorRT acceleration, with an average of 2.5ms per session.

When I get the image of the zed camera (np.array), and crop and merge it,（It will not be 1080p, depending on the people cropped out of the image, for example 1000x300）, the inference time is also normal.

But when I directly reasoned about the image of ZED (1080p), the preprocessing time became 10ms, which is much slower.

What could be the cause and how can I fix it?

The model I used：

model = YOLO("Models/yolo11n-seg.pt")
model.export(format="engine", dynamic=True, half = True, device=0)

Acquisition of ZED images：

bodies_status = zed.retrieve_bodies(bodies, body_runtime_param)
image_status = zed.retrieve_image(image, sl.VIEW.LEFT)

if bodies_status != sl.ERROR_CODE.SUCCESS or image_status != sl.ERROR_CODE.SUCCESS:
  continue

frame = image.get_data()

yolo inference

results = cutout_model.predict(
        img[:, :, :3],
        device="cuda:0",
        classes=[0],
        conf=0.4,
        retina_masks=True,
        verbose=True,)

[UltralyticsAssistant]

👋 Hello @IASUI, thank you for your interest in Ultralytics 🚀! We recommend a visit to the Docs for new users where you can find many Python and CLI usage examples. Many common questions may already be addressed there.

From your description, it seems like this might be related to the ZED camera image preprocessing. If this is a 🐛 Bug Report, we kindly ask that you provide a minimum reproducible example to help us investigate further. This typically includes a short, self-contained script with all relevant code and details to reproduce the issue (including input data, model configuration, and expected/observed behavior).

Additionally, to ensure the latest updates are applied and no prior fixes are missed, please confirm you’ve updated to the latest version of the ultralytics package, which you can do using the following command:

pip install -U ultralytics

Additional Context

For image preprocessing, you might want to double-check the image data type, format, and pixel values coming from the ZED camera. Differences in image formats (e.g., color channels, data type, or scaling) could potentially cause additional preprocessing overhead.

Join the Ultralytics community where it suits you best 🎧:

• For real-time chat, head to Discord
• Prefer in-depth discussions? Visit Discourse
• Dive into threads on our Subreddit

Environments

YOLO can be run in various verified environments for optimal performance, including:

• Notebooks with free GPUs:
• Google Cloud Deep Learning VM: See GCP Quickstart Guide
• Amazon Deep Learning AMI: See AWS Quickstart Guide
• Docker Image: See Docker Quickstart Guide

Status

If this badge is green, all Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLO Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

This is an automated response to help guide you, but no worries—a member of our Ultralytics engineering team will also assist you as soon as possible! 😊

[glenn-jocher]

The increase in preprocessing time for direct ZED 1080p images could be due to additional operations required to handle the raw image format, such as color conversion, resizing, or memory alignment. To optimize, ensure the frame format matches the expected input format of the model (e.g., RGB) and verify preprocessing steps are efficient. Profiling each step in the pipeline may help identify bottlenecks. For further insights, refer to the YOLO TensorRT documentation.

[Y-T-G]

What if you use img[..., :3].copy() in prediction?

[Y-T-G]

What is the value of print(img.shape)?

[IASUI]

(1080, 1920, 4)

[Y-T-G]

Maybe you should open an issue on the ZED community asking why it takes so long to copy.

[IASUI]

ok，thanks

[glenn-jocher]

You're welcome! If you have any further questions or need additional assistance with YOLO or Ultralytics, feel free to ask.