Deep Neural Networks Learn Meta-Structures from Noisy Labels in Semantic Segmentation

Introduction

This is the official code of

Deep Neural Networks Learn Meta-Structures from Noisy Labels in Semantic Segmentation (AAAI-2022 Oral).

Elucidating Meta-Structures of Noisy Labels in Semantic Segmentation by Deep Neural Networks (under review)

Data preparation

You need to download the ER and MITO datasets.

Your directory tree should be look like this:

$SEG_ROOT/datasets
├── er
│   ├── test
│   │   ├── images
│   │   ├── labels
│   ├── train
│   │   ├── images
│   │   ├── labels
│   ├── val
│   │   ├── images
│   │   ├── labels
├── mito
│   ├── test
│   │   ├── images
│   │   ├── labels
│   ├── train
│   │   ├── images
│   │   ├── labels
│   ├── val
│   │   ├── images
│   │   ├── labels
├── txt
│   ├── er
│   │   ├── train
│   │   │   ├── train_gt.txt
│   │   │   ├── train_noisyLabel.txt
│   │   └── val.txt
│   ├── mito
│   │   ├── train
│   │   │   ├── train_gt.txt
│   │   │   ├── train_noisyLabel.txt
│   │   └── val.txt

Training

To train model, you should save the datapath into a __.txt file and put it into the txt dictionary, then run main.py for training.

Testing

To test the segmentation performance, you should first run evaluation/inference.py to save the outputs of testing sets in train_log (Use parameters train_dir and test_ckpt_epoch to change the path of pre-trained models).

Then, you can run evaluation/inference.py to get different metrics scores such as IOU, F1 and others on testing set. (Use parameter test_data_dir to change the testing datapath __.txt. Use parameter prd_dir to change the saved predictions path of testing sets).

Reference

[1] Deep Neural Networks Learn Meta-Structures from Noisy Labels in Semantic Segmentation. Yaoru Luo, Guole Liu, Yuanhao Guo, Ge Yang Accepted by AAAI-22. download

Contributing

Code for this projects developped at CBMI Group (Computational Biology and Machine Intelligence Group).

CBMI at National Laboratory of Pattern Recognition, INSTITUTE OF AUTOMATION, CHINESE ACADEMY OF SCIENCES.

YaoruLuo / Meta-Structures-for-DNN