README.md 1.13 KB
Newer Older
ahmetkerem's avatar
ahmetkerem committed
1
# Label Noise Injection Tools
begumdemir's avatar
begumdemir committed
2
This page contains links to the tools developed at the [Remote Sensing Image Analysis group](https://www.rsim.tu-berlin.de/menue/remote_sensing_image_analysis_group/) to inject synthetic multi-label noise into image datasets. The obtained noisy labeled datasets can be used in the experiments to evaluate the robustness of the machine learning models against label noise. Both tools given below work on [Numpy](https://numpy.org/) arrays and are independent of specific deep learning libraries.
akakream's avatar
akakream committed
3

begumdemir's avatar
begumdemir committed
4
## 1) Multi Label Noise Injection Tool
tristan.kreuziger's avatar
tristan.kreuziger committed
5
[Multi-Label Noise Injection tool](https://git.tu-berlin.de/rsim/multi_label_noise) contains a set of helper functions in order to create noisy multi-label matrices. The tool allows to inject additive (extra classes) and subtractive (missing classes) noise separately.
akakream's avatar
akakream committed
6

akakream's avatar
akakream committed
7
8
![](images/multi-label_noise_injection_figure.png)

begumdemir's avatar
begumdemir committed
9
## 2) Noisifier Tool 
tristan.kreuziger's avatar
tristan.kreuziger committed
10
[Noisifier tool](https://git.tu-berlin.de/rsim/noisifier) allows to add synthetic multi-label noise to the multi-labeled image datasets. It is also applicable to the datasets containing single-labeled images.
akakream's avatar
akakream committed
11

tristan.kreuziger's avatar
tristan.kreuziger committed
12
![](images/noisifier_figure.jpg)