masterwu2115 / crow_tiling

This is the official implementation of the paper 'Tackling the Background Bias in Sparse Object Detection via Cropped Windows' Here you can find a Pytorch data set which reduces the background bias for sparse recordings and allows the usage of higher resolutions during training. The technique is only applicable the object detection.

Description

It is implemented as a data set wrapper (see CroWTiledDataset), which accepts a data set defining. This data set should implement the BaseDataset interface.

Example

1. Install the requirements.txt
2. Execute "python3 example.py"
3. output/ contains the created tiles. (All classes are visualized with green bounding boxes)

In the default configuration, 512x512 tiles are created. Further the full frame is added with a downscaling factor of 0.5. Empty tiles were discarded.

Folder structure

• 'crow_dataset.py' contains the data set wrapper. This is the only part necessary for CroW
• 'example.py' shows the usage of the tiling data set on an example MS COCO format data set
• 'example' contains an example dataset in MS COCO format
• 'output' will contain the created tiles with bounding boxes. (All classes are visualized with green bounding boxes)

Citation

@InProceedings{Varga_2021_ICCV,
    author    = {Varga, Leon Amadeus and Zell, Andreas},
    title     = {Tackling the Background Bias in Sparse Object Detection via Cropped Windows},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
    month     = {October},
    year      = {2021},
    pages     = {2768-2777}
}