Deformable-Sketch(Edge)-Detection-Network

Sketch detection network with deformable convolution.
When the input image is given, then model outs it's sketch(edge) version.
I use my custom handmake sketch(edge) dataset instead general benchmark datasets.

Test Envrionment

• NVIDIA A5000 24G
• Train 512 x 512 image
• Model params: 6.2M

Dataset

[Anime] Dataset (included nude pictures, and not multiscale)

• train: 120 images, edges
• val: 10 images, edges

Performance

Anime

Model	ODS↑	OIS↑	LPIPS(edge)↑	LPIPS(image)↑
UAED(pretrained by BSDS)	0.5417	0.5502	0.6500	0.5286
MuGE(pretrained by BSDS, α = 1.0)	0.5502	0.5721	0.6830	0.5465
DSDN(Anime)	0.6340	0.6389	0.7735	0.6323

BSDS500

Model	ODS↑	OIS↑	LPIPS(edge)↑	LPIPS(image)↑
UAED(pretrained by BSDS)	0.8410	0.8470	0.6519	0.3352
MuGE(pretrained by BSDS, α = 1.0)	0.850	0.8560	0.6899	0.3403
DSDN(Anime)	0.7354	0.7354	0.5301	0.4022

Results

Aru(BlueArchive)

Shiroko(BlueArchive)

Wakamo(BlueArchive)

Surtr(Arknights)

Amiya(Arknights)

Theresa(Arknights)

Benchmark(UAED, MuGE)

BSDS500

• the samples placed in the order UAED, MuGE, DSDN

Humans or Animals

• poor performance about humans or animals detection.(Noise sensitive & too many detect detail informations)

Structures

• good performance about structures than humans or animals detection.

Anime

Pictures(like above examples)

• good performance about pictures than sota edge detection networks.