• State
  • Locked Locked
  • Replies 3 replies
  • Subscribers 98 subscribers
  • Views 219 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

Original question:

TDA4VH-Q1: The model converted from Import and Edge AI are different

TDA4VM: I wonder if TDA4VM can achieve our goal of completing the entire process in less than 10ms.

Kuan-Hung Liu
Prodigy 100 points
Part Number: TDA4VM

Hi, we want to ensure that TDA4VM can achieve an end-to-end processing time of less than 10ms, which includes image pre-processing, model inference, and post-processing steps like tracking. The backbone model is currently ResNet18, and the input image size is 640 x 384. The main tasks are lane prediction and 3D object detection (3DOD). The final output size will be approximately 14N + 32000, depending on the number of detected objects (N). Do you have any recommendations for meeting these requirements?

Additionally, would using TIDL-RT for model inference be more beneficial compared to OSRT?

  • 0
    TI__Mastermind 24041 points

    Due to bandwidth, I cannot assess your question at this time.  Please expect a delay until 20/11/23.

    Thank you for your patience.

  • 0
    Prodigy 100 points

    Hi,
    Do you have any idea about my questions?

  • 0 in reply to Kuan-Hung Liu
    TI__Mastermind 24041 points

    Hi,

    Apologies for delay in my response due to bandwidth.

    Lets me explain OSRT flow vs TIDL-RT.

    You can you TIDL-RT when every layer of your model is supported from TIDL side. Where as in case of OSRT, unsupported layers will be offloaded to ARM core for execution, yes this will defiantly adds delay in terms of internal IPC and ARM (Scalar code execution) etc but end to end model inference is ensured.

    Based on your use case, here are few things that you can try.

    1. If all the layers are supported, you use TIDL-RT to generate the model artifacts

    2. Use TIDL-RT to do model import (Inference on target EVM) get the benchmark.

    3. Compare the benchmark result with your requirement.