Skip to content

Latest commit

 

History

History
112 lines (103 loc) · 3.37 KB

README.md

File metadata and controls

112 lines (103 loc) · 3.37 KB
Raw

Use Builtin Datasets

A dataset can be used by accessing DatasetCatalog for its data, or MetadataCatalog for its metadata (class names, etc). This document explains how to setup the builtin datasets so they can be used by the above APIs. Use Custom Datasets gives a deeper dive on how to use DatasetCatalog and MetadataCatalog, and how to add new datasets to them.

MGNet has builtin support for the Cityscapes and KITTI-Eigen dataset. The datasets are assumed to exist in a directory specified by the environment variable DETECTRON2_DATASETS. Under this directory, MGNet will look for datasets in the structure described below, if needed.

$DETECTRON2_DATASETS/
  cityscapes/
  kitti_eigen/

You can set the location for builtin datasets by export DETECTRON2_DATASETS=/path/to/datasets. If left unset, the default is ./datasets relative to your current working directory.

Cityscapes

Go to the # Cityscapes website and download the following dataset parts (You have to create an account to be able to download the zip files):

leftImg8bit_trainvaltest.zip
gtFine_trainvaltest.zip
disparity_trainvaltest.zip
camera_trainvaltest.zip
leftImg8bit_sequence_trainvaltest.zip
disparity_sequence_trainvaltest.zip

Extract all files and generate the panoptic dataset using the preparation script python prepare_cityscapes.py.

Expected dataset structure for Cityscapes:

cityscapes/
  camera/
    train/
      aachen/
        camera.json
    val/
    test/
  disparity/
    train/
      aachen/
        disparity.png
    val/
    test/
  gtFine/
    train/
      aachen/
        color.png, instanceIds.png, labelIds.png, polygons.json,
        labelTrainIds.png
      ...
    val/
    test/
    # below are generated Cityscapes panoptic annotation
    cityscapes_panoptic_train.json
    cityscapes_panoptic_train/
    cityscapes_panoptic_val.json
    cityscapes_panoptic_val/
    cityscapes_panoptic_test.json
    cityscapes_panoptic_test/
  leftImg8bit/
    train/
    val/
    test/
  leftImg8bit_sequence/
    train/

KITTI-Eigen

Please follow the instructions in packnet-sfm to download the KITTI-Eigen dataset in the correct structure. Note that since KITTI-Eigen does not provide panoptic annotations, MGNet training on KITTI-Eigen requires pseudo_labels using tools/generate_pseudo_labels.py with a pretrained Cityscapes model. See getting started for instructions on how to train a model on KITTI-Eigen.

Expected dataset structure for KITTI:

kitti_eigen/
  2011_09_26/
    2011_09_26_drive_0001_sync/
      image_02/
        data/
          0000000000.png
          ...
        cam.txt
        poses.txt
        timestamps.txt
      image_03/
      oxts/
      proj_depth/
        groundtruth/
          image_02/
            0000000005.png
            ...
          image_03/
        velodyne/
    2011_09_26_drive_0002_sync/
    ...
    calib_cam_to_cam.txt
    calib_imu_to_velo.txt
    calib_velo_to_cam.txt
  2011_09_28/
  2011_09_29/
  2011_09_30/
  2011_10_03/
  data_splits/
    eigen_test_files.txt
    eigen_zhou_files.txt
    ...