• State TI Thinks Resolved
  • Locked Locked
  • Replies 2 replies
  • Answers 2 answers
  • Subscribers 97 subscribers
  • Views 782 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.

object detection training tool question

Josh HUANG
Intellectual 290 points

Dear all:    

     I have a question about TI’s Object Detection training toolbox of Acf-jacinto.

     I use the training dataset from Caltech - INRIA Pedestrian Test Dataset,

     and input database into Piotr's toolbox(TI fork source) and TI Acf-jacinto toolbox.

    We obtain different results, Piotr's toolbox detect all pedestrian in first test image but TI Acf-jacinto toolbox only detect 3 pedestrian.

    The miss rates are not the same (Piotr's is 16%,  TI Acf-jacinto is 35% as attached).

    Is this correct detect result & miss rate for TI Acf-jacinto toolbox?

    

  • 0
    TI__Mastermind 42515 points
    Hi Josh,

    I have forwarded your question to algorithms experts.

    Regards,
    Yordan
  • 0
    TI__Genius 11546 points
    I think the pictures that you showed are correct. However, there is an aspect of input resolution. The choice is between running a slow algorithm and hence using a small resolution image as input or running a fast algorithm and using a large resolution image as input.

    The algorithm in acf-jacinto training tool is tuned to be a fast variant of the algorithm in Piotr's toolbox. You can run the object detection at much higher resolution using this algorithm. We have seen that doing object detection on a higher resolution improves the detection quality. Hence we have used some trade-offs to increase the speed/input resolution of the algorithm.