Datasets

2D Datasets

1. and 2. Reflection and Rotation Symmetry Competition

A. Reflection [link]

Reflection (Single) : 100 Training Images
Reflection (Multiple) : 100 Training Images

B. Rotation [link]

Rotation (Single) : 100 Training Images
Rotation (Multiple) : 100 Training Images

C. Sym-COCO [ref link] [rot link]

250 training images are selected from MS-COCO
Human labeled reflection symmetry axis (two points) and rotation symmetry centers (one point) are used as basis for ground truth symmetries.

3. Translation 1 (frieze)

Frieze Dataset [link]

50 Training images
Frieze Patterns in real images

4. Translation 2 (wallpaper)

Wallpaper Dataset [link]

100 Training images
Wallpaper Patterns in real images

5. Skeleton Symmetry (medial-axis) Competition

SK-LARGE [link]

1491 cropped images from MS-COCO (746 train, 245 validation, 500 testing).
Object skeletons for various categories, such as: person, horse, airplane etc.
Medial point location and scale.

BMAX500 [link]

500 images (200 train, 100 val, 200 test) from the Berkeley Segmentation Dataset.
Medial axes of human-annotated image segments.
5-7 annotations for each image.
Medial point location and scale.

3D Datasets

All 3D datasets and annotations have the same format, and all submissions must respect the annotation format, as described here.

The training datasets were released on May 15, 2017 and the test data sets were released May 29, 2017.

3D Symmetry Detection

The goal is to identify the major symmetries for each object. For the global symmetry dataset the reflections are valid for the whole scene, while for the local symmetries the symmetry support is a region of limited size. A secondary goal is to also estimate the symmetry support region in form of a object aligned bounding box.

Note that the symmetries might not always be perfect and we allow for small deviations, e.g. the exhaust pipes and the steering wheel of a car are only on one side of the car while all other components are perfectly symmetric.

1. Global Reflection Symmetry (Synthetic Data)

3D-globalSym-synth [sample][training][test]

1354 training/441 test models.
Up to 3 symmetry planes per object.

2. Local Reflection Symmetry (Synthetic Data)

3D-localSym-synth [sample][training][script][test]

200 training/200 test 3D scenes.
Scenes automatically generated by randomly placing different objects on a flat surface.
Script to generate more scenes using the models from the 3D-globalSym-synth Dataset.

3. Global Reflection Symmetry (Real Data)

3D-globalSym-real [sample][training][test]

20 training/20 test 3D models.
Symmetry planes for different objects: chair, buildings, indoor spaces.
Up to 2 reflective symmetries per scene.

4. Local Reflection Symmetry (Real Data)

3D-localSym-real [sample][training][test]

21 training/21 test 3D scenes.
Symmetry planes and bounding boxes for all symmetric objects in the scene.
Up to 10 objects per scene.

Acknowledgements and Data Usage

3D Synthetic datasets: These models originate from the Archive3D database and have been annotated by us. Please do not redistribute the data and respect their terms and conditions (currently, free usage).

For the local symmetry dataset, the models are randomly selected and composed into a scene with multiple objects.

3D Real datasets: The datasets are manually labeled with ground truth symmetries. The models originate from the following publicly available datasets:
- SUN3D scenes and ICL-NUIM sequences [Link] from the paper :

Sungjoon Choi, Qian-Yi Zhou, Vladlen Koltun, "Robust Reconstruction of Indoor Scenes", CVPR 2015

Object Scans [Link] from the paper:

Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, Vladlen Koltun, "A Large Dataset of Object Scans", Technical Report, arXiv:1602.02481, 2016

Please cite the corresponding papers if you use these datasets.

If you use any of the labeled or generated local symmetry data, please also cite the corresponding workshop paper:

Christopher Funk, Seungkyu Lee, Martin R. Oswald, Stavros Tsogkas, Wei Shen, Andrea Cohen, Sven Dickinson and Yanxi Liu. "2017 ICCV Challenge: Detecting Symmetry in the Wild", IEEE International Conference on Computer Vision (ICCV) Workshop, 2017

Google Sites

Report abuse