TGRNet-REP

This is the official code for TGRNet-REP.

Requirements

python==3.8

einops==0.3.0

cryptography==43.0.0

torch==2.3.0

numpy==1.24.4

pandas==2.0.3

opencv-python==4.4.0.46

Visualization

Visualization on WHU-Stereo dataset

(a) left image. (b) disparity ground truth. (c) SGM. (d) PSMNet. (e) HMSMNet. (f) RAFT. (g) DLNR. (h) Ours.

Visualization on US3D dataset

(a) left image. (b) disparity ground truth. (c) SGM. (d) PSMNet. (e) HMSMNet. (f) RAFT. (g) DLNR. (h) Ours.

Dataset Preparation

• WHU-Stereo

• US3D

  Note：If you want to use our preprocessed US3D dataset, please download：

Pretrained weights

The pretrained weights for WHU-Stereo and US3D datasets are available at:

https://pan.baidu.com/s/1uuZl9WvSpLw6wgNkq9-kTA password: 4k7j

Please place them in the main directory.

Evaluation on WHU-Stereo

python evaluate_WHU.py --test_left_dir dataset/WHU-Stereo/with_GT/test_all/left --test_right_dir dataset/WHU-Stereo/test_all/right --test_disp_dir dataset/WHU-Stereo/test_all/disp --test_save_path results/whu --device cuda

test_left_dir: Directory containing the left image

test_right_dir: Directory containing the right image

test_disp_dir: Directory containing the disparity ground truth

The results will be:

EPE: 1.586  D1:12.63%

Comparsion with other methods

The best score for each metric is marked in bold.

Method	EPE(Px)	D1(%)
SGM	4.989	36.22
PSMNet	2.183	21.95
HMSMNet	2.040	19.00
RAFT	1.729	14.12
DLNR	1.864	16.56
TGRNet(ours)	1.586	12.63

Evaluation on US3D

python evaluate_US3D.py --test_left_dir dataset/US3D/test_all/left --test_right_dir dataset/US3D/test_all/left --test_disp_dir dataset/US3D/test_all/left --test_save_path results/us3d --device cuda

test_left_dir: Directory containing the left image

test_right_dir: Directory containing the right image

test_disp_dir: Directory containing the disparity ground truth

The results will be:

EPE: 1.313  D1:7.17%

Comparsion with other methods

The best score for each metric is marked in bold.

Method	EPE(Px)	D1(%)
SGM	2.398	19.93
PSMNet	1.499	9.22
HMSMNet	1.473	9.17
RAFT	1.366	7.72
DLNR	1.389	8.03
TGRNet(ours)	1.313	7.17

Test on your own image!

The following command is provided to allow you to test your own dataset! We give an example:

python evaluate_single.py --left_path KM_left_0.tiff --right_path KM_right_0.tiff --save_path KM_pred_0.tiff --mode 16bit device cuda

mode: 16bit(single channel) or 8bit(RGB 3 channels)

The result will be saved at 'KM_pred_0.tiff'!

Good luck!

Thank you for using our code!