The WiderPerson dataset is a pedestrian detection benchmark dataset in the wild, of which images are selected from a wide range of scenarios, no longer limited to the traffic scenario.
We choose 13,382 images and label about 400,000 annotations with various kinds of occlusions. We randomly select 8000/1000/4382 images as training, validation and testing subsets. Similar to CityPersons and WIDER FACE datasets, we do not release the bounding box ground truths for the test images. Users are required to submit final prediction files, which we shall proceed to evaluate.
More information can be found by reading WiderPerson: A Diverse Dataset for Dense Pedestrian Detection in the Wild.
Each image of training and valiadation subsets in the "./Images" folder (e.g., 000001.jpg) has a corresponding annotation text file in the "./Annotations" folder (e.g., 000001.jpg.txt). The annotation file structure is in the following format:
...
< number of annotations in this image = N >
< anno 1 >
< anno 2 >
...
< anno N >
...
where one object instance per row is [class_label, x1, y1, x2, y2], and the class label definition is:
...
< class_label =1: pedestrians >
< class_label =2: riders >
< class_label =3: partially-visible persons >
< class_label =4: ignore regions >
< class_label =5: crowd >
...
The dataset can be downloaded either through Google drive or Baidu drive.
No associated challenges have been found for this dataset.
The dataset is lincenced under a non-commercial licence.