BiCoD

Code for our paper :
BiCoD: Bidirectional Correlated Depth Guidance for Domain Adaptation Segmentation

1. Installation

Please create and activate the following conda envrionment. To reproduce our results, please kindly create and use this environment.

# create environment
conda update BiCoD
conda env create -f environment.yml
conda activate BiCoD

Code was tested on an NVIDIA 3090Ti with 24G Memory.

Our results

2. Datasets

CITYSCAPES: Follow the instructions in Cityscape to download the images and validation ground-truths. Please follow the dataset directory structure:

<CITYSCAPES_DIR>/             % Cityscapes dataset root
├── leftImg8bit_trainvaltest/ % input image (leftImg8bit_trainvaltest.zip)
├── depth/                    # From https://people.ee.ethz.ch/~csakarid/SFSU_synthetic/, also downloadable at https://www.qin.ee/depth/
├── disparity/                % stereo depth (disparity_trainvaltest.zip)
└── gtFine_trainvaltest/      % semantic segmentation labels (gtFine_trainvaltest.zip)

SYNTHIA: Follow the instructions here to download the images from the SYNTHIA-RAND-CITYSCAPES (CVPR16) split. Download the segmentation labels from CTRL-UDA using the link here. Please follow the dataset directory structure:
```
<SYNTHIA_DIR>/                % Synthia dataset root
├── RGB/                      % input images
├── GT/                       % semseg labels labels
└── Depth/                    % depth labels
```

GTA5: Follow the instructions here to download the images from the SYNTHIA-RAND-CITYSCAPES (CVPR16) split. Download the segmentation labels from CTRL-UDA using the link here. Please follow the dataset directory structure:

<GTA_DIR>/                    % Synthia dataset root
├── images/                   % input images
├── labels/                   % semseg labels labels
└── disparity/                % Our generated monodepth 0-65535 in normalized disparity, downloadable at https://www.qin.ee/depth/

3.Train a BiCoD model

comming soon

4. Test a trained BiCoD model

Pre-trained models are provided (Google Drive).

# Test the model for the SYNTHIA2Cityscapes task
python3 evaluateUDA.py --full-resolution -m deeplabv2_synthia --model-path=<"model path">
# Test the model for the GTA2Cityscapes task
python3 evaluateUDA.py --full-resolution -m deeplabv2_gta --model-path=<"model path">

5. Results

Reported Results on SYNTHIA2Cityscapes (The reported results are based on 5 runs instead of the best run.)

Method	mIoU*(13)	mIoU(16)
CBST	48.9	42.6
FDA	52.5	-
DADA	49.8	42.6
CTRL	51.5	45
CorDA	62.8	55.0
BiCoD	64.4	56.8

Acknowledgement

This codebase depends on CorDA, AdaBoost_seg and DenseMTL. Thank you for the work you've done!!!
DACS is used as our codebase official
SFSU as the source of stereo Cityscapes depth estimation Official

Data links

Those compents are provided by CorDA,Thanks!

Download links
- Stereo Depth Estimation for Cityscapes
- Mono Depth Estimation for GTA
- SYNTHIA Depth and images SYNTHIA-RAND-CITYSCAPES (CVPR16)

yuheyuan/BiCoD