Distance-dependent anomaly detection

This repo enables measuring the sensitivity of anomaly detection to distance from the camera on the LostAndFound dataset as described in Dense anomaly detection by robust learning on synthetic negative data.

Requirements

Available in requirements.txt

• Pillow
• prettytable
• torch
• torchvision

Setup

For LostAndFound dataset download simply run:

./prepare_dataset.sh

Usage

Simple demo evaluation script:

python evaluate.py

Example output:

+----------------+-------+-------+-------+-------+-------+-------+-------+-------+-------+
|   Range (m)    |  5-10 | 10-15 | 15-20 | 20-25 | 25-30 | 30-35 | 35-40 | 40-45 | 45-50 |
+----------------+-------+-------+-------+-------+-------+-------+-------+-------+-------+
|       AP       |  48.3 | 52.94 | 55.21 | 54.06 | 51.94 | 42.07 | 37.11 | 43.75 | 35.23 |
| FPR at TPR 95% |  7.95 | 10.23 | 11.33 | 16.42 | 20.52 | 26.14 | 28.98 | 34.32 |  43.8 |
|     AUROC      | 98.07 |  97.7 | 97.41 | 96.52 | 95.36 |  93.5 | 91.45 |  90.2 | 86.99 |
+----------------+-------+-------+-------+-------+-------+-------+-------+-------+-------+

If used, please cite:

@article{grcic22arxiv,
  author    = {Matej Grcic and
               Petra Bevandic and
               Zoran Kalafatic and
               Sinisa Segvic},
  title     = {Dense anomaly detection by robust learning on synthetic negative data},
  journal   = {CoRR},
  volume    = {abs/2112.12833},
  year      = {2021}
  }