This code is official code of MS-UDA: Multi-Spectral Unsupervised Domain Adaptation for Thermal Image Semantic Segmentation.
If you want to use KP dataset, go to here And, Please cite our project.
The KP dataset structure is:
-day2night
-labels
-val_day
-val_night
-visualize
-pseudo_KP
- day
- night
- val_day
- val_night
day2night
This folder contains fake nighttime thermal images generated by Cycle-GAN, trained with 3283 daytime images and 3095 nighttime images.
labels
This folder contains real label for KAIST Pedestrian segmentation dataset. The folder 'val_day' contains the daytime 503 rgb-t images which is used to train the supervised network for comparison. The folder 'val_night' contains the nighttime 447 rgb-t images which is used to validate MS-UDA. The folder 'visualize' contains visualization of the label
pseudo_KP
This folder contains the rgb-t input images and their pseudo-labels. The folder 'day' contains the 3283 daytime rgb-t images for domain adaptation and day-to-night translation. The folder 'night' contains the 3095 nighttime rgb-t images for day-to-night translation. The folder 'val_day' contains the 503 rgb-t daytime images. The folder 'val_night' contains the 447 rgb-t nighttime images.
This folder contains the filename for KP dataset.
It contains:
-day_rgb/th.txt
-night_rgb/th.txt
-val_day_rgb/th.txt
-val_night_rgb/th.txt
day_rgb/th.txt has the name of 3283 daytime rgb/thermal images for UDA and day-to-night translation.
night_rgb/th.txt has the name of 3095 nighttime rgb/thermal images which is used for day-night translation.
val_day_rgb/th.txt has the name of 503 daytime rgb/thermal images for supervised learning method to compare the results with UDA.
val_night_rgb/th.txt has the name of 447 nighttime rgb/thermal images for validation.
'..pred.png' is the visualized pseudo-labels. '..pseudo.png' are pseudo labels. '...rgb/th.png' are the input images.