search

suojiashun / HIT-UAV-Infrared-Thermal-Dataset

A high-altitude infrared thermal dataset for Unmanned Aerial Vehicle-based object detection
Creative Commons Attribution 4.0 International
126 stars 21 forks source link

Entire dataset is mislabeled by an offset #6

Open andrewjong opened 1 year ago

andrewjong commented 1 year ago

Hi,

Thanks for your hard work to release this dataset!

I found a minor issue that should be easily fixable. It seems the way your data is formatted doesn't match the way your Dataset class reads in the data.

The HITUAVDataset expects the data to be [center x, center y, width, height]. (This matches Standard YOLO format after normalization and is the correct assumption for YOLO format.)

However, your data is actually labeled [min x, min y, width, height] (a mix between Pascal VOC and YOLO format; true Pascal VOC format would be [min x, min y, max x, max y]).

Therefore, as can be seen when visualizing the ground truth labels, all the boxes are incorrectly offset by -w/2, -h/2.

Can the data be fixed to match a standard format, e.g. YOLO? Thanks!

I've only examined the standard annotation, haven't looked at the oriented bboxes.

Examples: image image image image

I have verified this is not simply a visualization error, as my visualization code performs correctly on other data in YOLO format.

Otherwise it's a very nice dataset.

suojiashun commented 1 year ago

Thank you for your attention! For the standard label with VOC format, the folder _normalxml has been provided. For the label in _normalxml, the objects are labeled with [xmin, ymin, xmax, ymax], which is standard Pascal VOC format. Additionally, we provide a demo tool, which can convert VOC to YOLO format, in the folder tools/output/voc2yolo.py.

Hope we can help you!