Dataset preparation

April 1, 2022 ยท View on GitHub

To train the model, data annotations should be converted to COCO format and arranged in the following format:

${DATA_PATH}
|-- annotations
    |-- train_semantic_line.json
    |-- val_semantic_line.json
|-- images
    |-- train_semantic_line
    |-- train_semantic_line_mag
    |-- val_semantic_line
    |-- val_semantic_line_mag

where train_semantic_line / val_semantic_line folders contain the images used for training and validating.

[Optional] train_semantic_line_mag / val_semantic_line_mag folders contain the filtered gradient magnitude images processed by selected edge operators (such as Scharr or Sobel) used for Magnitude Loss.

Currently we have 5 KAIST_URBAN sequences sampled (1 selected from every other 5 images) and labeled:

Data set# Labeled
KAIST seq261946
KAIST seq291479
KAIST seq304227
KAIST seq384318
KAIST seq393729
KITTI14999

Splited train/val images and annotations in COCO format as well as .xml format can be found:

KAIST URBAN: Labels

As for the images, please refer to the official website. We use the stereo_left images.

KITTI: Images / Labels

To generate image gradient magnitudes

Check src/_get_gradient_magnitude_images.py