DL model performances are not same in Mech-DLK and Mech-Vision

Software version:

  • Mech-DLK 2.3.0
  • Mech-Vision 1.7.1
  1. Application background description:

    Application is typical picking scenario for standard metal parts. However, due to the slip sheet between each layer and the metal attached to each other closely, instance segmentation based on deep learning is necessary to improve the vision project robustness. The responsibility for DL model is to distinguish the metal parts and slip sheet and to segment the metal parts.

  2. Problem:

    After training and verification, the DL model could segment the parts correctly. However, when the DL model implement in Mech-Vision, the result confidence value is not satisfied. Moreover, the edge of segmentation result is not guaranteed.

  • The verification result in Mech-DLK is as follows:

  • The result based on the same DL model and the same image in Mech-Vision is as follows:

  1. Question:
  • What will cause such problems?
  • How to figure out the reason? Any instructions for solving such problems?

Please check if the ROI set in Mech-Vision is consistent with that in Mech-DLK. From the image, it can be seen that the ROI for training/validation in Mech-DLK is the full image. In Mech-Vision, the default ROI is not used, but the user has defined their own ROI.