BiconNets

June 16, 2023 ยท View on GitHub

Ziyun Yang, Somayyeh Soltanian-Zadeh and Sina Farsiu

Recently accepted by Pattern Recognition.

Paper at: https://arxiv.org/abs/2103.00334

BiconNet also achieved great improvements on medical segmentation and multi-class, please visit BiconNet-Medical for details

Please also check our latest DconnNet paper on CVPR 2023

Updates: 06/15/2023: Major bugs found and fixed. The codes are now able to run directly.

Requirement: Pytorch 1.7.1

This code including three parts:

  1. Codes for customizing BiconNet wtih other backbones (/general)
  2. Codes for reproducing the paper results (/paper_result)
  3. Evaluation Code (/evaluation)

Customize the BiconNet based on your own network. (/general)

If you want to construct the BiconNet based on your own network, there are four simple steps:

  1. replace your network's one-channel output fully connected layers with 8-channel FC layers.

For training:

  1. generate the ground truth connectivity masks using the function 'sal2conn' in utils_bicon.py

  2. replace your own loss function with Bicon_loss: you can edit the connect_loss.py

For testing:

  1. use the function 'bv_test' in utils_bicon.py after you get the 8-channel connectivity map output to get your final saliency prediction.

Reproduce the results in the paper (/paper_result)

For traing: cd /MODEL_NAME/bicon/train python train.py

For testing: cd /MODEL_NAME/bicon/test python test.py

The pretrained models and maps can be downloaded from Google Drive

Results evaluation (/evaluation)

We use Matlab to evaluate the output saliency maps as did in: https://github.com/JosephChenHub/GCPANet

Citation

If you find this work useful in your research, please consider citing:

"Z. Yang, S. Soltanian-Zadeh, and S. Farsiu, "BiconNet: An Edge-preserved Connectivity-based Approach for Salient Object Detection", Pattern Recognition 121, 108231 (2022)"