search

tangsanli5201 / DeepPCB

A PCB defect dataset.
MIT License
358 stars 97 forks source link

About speed #4

Closed Apostatee closed 5 years ago

Apostatee commented 5 years ago

This paper says SSD achieves 64fps with input_size=512, but original paper of SSD only reaches 25fps with input_size=512, Could you please explain it in detail ?

tangsanli5201 commented 5 years ago

Hi, this is a good question since the efficiency is what we mainly concern for the ONLINE PCB defect detector. There is a mistake in your question that SSD achieves 64fps with input_size=640. (512 is the training size). There are following reasons:

  1. Smaller backbone structure: VGG-7 is used as the SSD backbone in PCB defect detection task, whose model size is only around 2MB. The original SSD use ResNet50 or more complex backbone structure.
  2. The PCB image is in grayscale but the input image for the original SSD is in RGB. Hope above instructions be useful.
Apostatee commented 5 years ago

Thanks tangsanli : )