TI Deep Learning Product User Guide
TIDL-RT: Output Post Processing in Test Application

Table of Contents

Introduction

TIDL-RT supports following post processing types in TIDL-RT test application to sanity test the model deployments.

  • postProcType - 1 : Classification Accuracy
  • postProcType - 2 : Object Detection Box drawing
  • postProcType - 3 : Pixel Level Color Blending
  • postProcType - 4 : Pixel Level MeanIOU Accuracy

Note: For all the post processing TIDL-RT import/infer config parameter "inFileFormat" shall be 2. All the these post processing are available as part of TIDL-RT test application and are only meant for demo purpose and are not of production quality.

Classification Accuracy

This post processing option can be enabled by setting postProcType = 1 in TIDL-RT import/infer configuration file. This option implements Top-1 and Top-5 accuracy computation for classification networks. The inData parameter in import/infer config file shall contain the list of all the input images along with their expected class index (ground truth). Please refer the imageNet_sample_val.txt file in ti_dl/test/testvecs/config folder as a reference to understand how to provide the input.

Following figure shows an example of how the output with this post processing option will be displayed:

Postproc_1.PNG
Classification Accuracy

In this example :

  • The fist number after "A:" is expected class index.
  • The Second is Top 1 Accuracy (1 means 100%) till current frame processing.
  • The Third is Top 5 Accuracy (1 means 100%) till current frame processing.
  • The Last is Predicted Class index for the current Frame

Object Detection Box drawing

This post processing option can be enabled by setting postProcType = 2 in TIDL-RT import/infer configuration file. This option generates an output image where a bounding box is drawn on all the detected objects on top of original input image. Each class is represented by a unique color. The input file shall contain the list of input images to be processed. The maximum number of classes in this post processing option is limited to 21. Note that there is no limit on number of detected objects.

Postproc_2.PNG
Object Detection BBox

Pixel Level Color Blending

This post processing option can be enabled by setting postProcType = 3 in TIDL-RT import/infer configuration file. This option generates an output image which is a blended image of input image with a unique color for each class index or depth estimation for every input pixel.

Postproc_3.PNG
Pixel Level Semantic Segmentation

Pixel Level MeanIOU Accuracy

This post processing option can be enabled by setting postProcType = 4 in TIDL-RT import/infer configuration file. This option will enable MeanIOU accuracy computation for pixel level segmentation networks. The input list file shall contain the list of input images with ground truth image for each input. The inData parameter in import/infer config file shall contain the list of all the input images along with their expected ground truth image. Please refer the segmentation_list.txt file in ti_dl/test/testvecs/config folder as a reference to understand how to provide the input. When Post Processing 4 is enabled, then pixel level color blending as mentioned in postProcType = 3 is enabled by default.