Benchmark dataset setup

In order to work our default config files, all data is expected to be in the datasets directory of this repository in the following directory structure:

aldi/
    datasets/
        cityscapes/
            leftImg8bit/
            leftImg8bit_foggy/
            annotations/
                cityscapes_train_instances.json
                ...
        sim10k/
            images/
            coco_car_annotations.json
        cfc/
            images/
                cfc_train/
                cfc_val/
                cfc_channel_train/
                cfc_channel_test/
            coco_labels/
                cfc_train.json
                ...

Cityscapes → Foggy Cityscapes

Images: Downloading Cityscapes and Foggy Cityscapes requires creating an account on the Cityscapes website. This is a multi-step process...

Create an account on the Cityscapes website and wait to be approved. This will be easier if you have an academic email address.
Install the CityScapesScripts command line utilities.
Use csDownload on the command line to download the leftImg8bit and leftImg8bit_foggy images, and place them in aldi/datasets/cityscapes/ as shown above.

Labels: For reproducibility, we provide JSON files containing the ground-truth bounding boxes we used for training and evaluation. As noted in our paper, these have differed in past codebases. Our files match the on-the-fly conversion done by Detectron2. See our code here for reference.

Cityscapes train labels

Cityscapes val labels

Foggy Cityscapes train labels

Foggy Cityscapes val labels

Sim10k → Cityscapes

Sim10k images: Download the Sim10k images here, and place them in aldi/datasets/sim10k/images/ as shown above.

Cityscapes images: Follow instructions for setting up Cityscapes above. Note you will only need the leftImg8bit images, not the foggy ones

Labels: For DAOD, Sim10k → Cityscapes is typically a single-class challenge consisting only of the "car" class. We provide labels postprocessed for this task here: