Skip to content

Latest commit

 

History

History
44 lines (31 loc) · 1.71 KB

summary_comparison_with_backbones.md

File metadata and controls

44 lines (31 loc) · 1.71 KB
Raw

Comparison of backbone networks

This directory contains experiment results of different backbones on MVTec AD dataset.

Purpose

It's quite easy to swap the backborn network in the PatchCore algorithm (default: Wide ResNet50 x2). It's meaningful to find a good backbone network which shows a good performance-speed tradeoff from an application view point.

What we've done

Try several backbone networks and evaluate their average image/pixel-level scores in the MVTecAD dataset.

Test Environment

  • CPU: Intel(R) Xeon(R) CPU E5-2690 v4 @ 2.60GHz (56 cores)
  • RAM: 128 GB

Conclution

The smaller ResNet (ResNet18, ResNet34) shows enough good scores even for their small computational cost. On the otherhand, very deep ResNet (ResNet101, ResNet152) shows a lower performance than ResNet50. Current tentative hypothesis is that the features used in the PatchCore algorithm are too deep (too far from input) and don't have enough high-resolution (raw) input information in them. In other words, we should add more shallower fatures in the case of very deep neural networks like ResNet101 and ResNet152 for exceeding ResNet50's score.