https://arxiv.org/pdf/2111.09099.pdf

Why

https://github.com/ristea/sspcab: v1 SSPCAB

1. simple & runnable code & Ranked #6 on Anomaly Detection on CUHK Avenue (TBDC metric), #8 on https://paperswithcode.com/dataset/shanghaitech

https://github.com/ristea/ssmctb: v2 SSMCTB

Ranked #4 on Anomaly Detection on CUHK Avenue (TBDC metric), #7 on https://paperswithcode.com/dataset/shanghaitech

Introduction

masked convolution (⇒activation maps) + CA (channel attention)

no SA: SA < SA + CA < CA for the data sets used, see Table 1.

Lsspcab: mean squared error between input & output maps.

v2: Self-Supervised Masked Convolutional Transformer Block for Anomaly Detection

M: 1x1xc; Ki: k’xk’xc;

only 2 hyper parameters: k’, d

{1, 2, 3} for k‘, and values in {0, 1, 2} for d

⇒k’=d=1 is the best, see table 1.

2 Related works

2.1 reconstruction-based approaches