#Dataset: [BaiduNetdisk][Google]
Hazy-JHU： Samples in this Hazy-JHU dataset are selected from JHU-CROWD++，whose whose weather conditions label is fog/haze. This hazy-weather dataset belongs to the real-world one, and the train-validation-test is still following the original spilt of JHU-CROWD++. There are 80 samples, 23 samples and 50 samples in training, validation and testing part of this dataset.
Hazy-ShanghaiTechRGBD： Samples in this Hazy-ShanghaiTechRGBD dataset are synthesized on ShanghaiTechRGBD dataset,accord- ing to the haze simulation algorithm.This hazy-weather dataset belongs to the synthetic one,, and the train-test is still following ShanghaiTechRGBD dataset. train_data contains images, train_depth and train_gt. (1) images: .jpg, 19201080 resolution. (2) train_gt: *.mat, the position of a head is annotated in point (x, y), where 0 <= x < 1920 and 0 <= y <= 1080. (43) train_bbox: pseudo box generated from point annotation and depth.
test_data contains imagesand test_bbox_anno. (1) The formats of images are similar to images in the train. (2) test_bbox_anno: the position of a head is annotated in bounding box (x1, y1, x2, y2), where (x1, y1) and (x2, y2) correspond to the coordinates of top-left corner and bottom-right corner.
If you find Hazy-ShanghaiTechRGBD and Hazy-JHU datasets useful, please consider cite our following work. @article{Kong2023HazyCC, title={Direction-aware attention aggregation for single-stage hazy-weather crowd counting}, author={Weihang Kong, Jienan Shen , He Li, Jiayu Liu, Junge Zhang}, journal={Expert Systems with Applications}, year={2023}, volume={225} pages={120088} }