Skip to content

Latest commit

 

History

History
45 lines (41 loc) · 3.08 KB

datagen.md

File metadata and controls

45 lines (41 loc) · 3.08 KB
Raw

🚷 Traffic sign templates

A subset of the near complete list of German traffic sign(s) is of interest to us. More specifically these. These signs form a subset (signs of interest) which formalise turn-restrictions in OSM. Using the templates for the signs of interest we build a synthetic training set following the idea presented in IJCN2019 and CVPR 2016.

🏭 Generate synthetic data

Use generate-synthetic-dataset to generate a synthetic sign dataset:

optional arguments:
  -h, --help            show this help message and exit
  --backgrounds BACKGROUNDS
                        Path to the directory containing background images to
                        be used (or file list).
  --templates-path TEMPLATES_PATH
                        Path (or file list) of templates.
  --augmentations AUGMENTATIONS
                        Path to augmentation configuration file.
  --distractors-path DISTRACTORS_PATH
                        Path (or file list) of distractors.
  --random-distractors RANDOM_DISTRACTORS
                        Generate this many random distractors for each
                        template.
  --out-path OUT_PATH   Path to the directory to save the generated images to.
  --max-images MAX_IMAGES
                        Number of images to be generated.
  --n JOBS              Maximum number of parallel processes.
  --max-template-size MAX_TEMPLATE_SIZE
                        Maximum template size.
  --min-template-size MIN_TEMPLATE_SIZE
                        Minimum template size.
  --background-size BACKGROUND_SIZE
                        If not None (or empty string), image shape
                        'height, width'

The following example generates a dataset of 2M images. The file augmentations.yaml specifies augmentation parameters (geometric template distortion, blending methods, etc.). Refer to the documentation of generate_task_args() in synthetic_signs.dataset_generator for a parameter description.

generate-synthetic-dataset --backgrounds=synthetic_signs/external/lists/Building_without_signs.list \
                           --templates-path=synthetic_signs/templates \
                           --out-path=experiments/synthetic-signfeld-dataset \
                           --n=16 \
                           --max-images=200000 \
                           --augmentations=resources/configs/augmentations.yaml

Sample training images can be found here