Sketch Classification

A PyTorch Implementation for Sketch Classification Networks.

Model Configuration

• Optimizer
  • Adam

DataSet

TU-Berlin sketch dataset

Model	input_size
(raw size)*	1111 * 1111
AlexNet	224 * 224
SketchANet	225 * 225
ResNet18	224 * 224
ResNet34	224 * 224
ResNet50	224 * 224
DenseNet121	224 * 224
Inception_v3	299 * 299

Model Parameters

Model	lr	clip_grad_norm(max_norm)	learning rate decay	weight_decay
AlexNet(pretrained)	2e-4	--	20	0.0005
AlexNet(scratch)	2e-5	0.5 - 100.0	30	0.0005
SketchANet(DogsCats)*	2e-5	0.5 - 1.0	30	0.0005
SketchANet(scratch)	2e-5	0.5 - 100.0	800	0.0001 - 0.0003
ResNet18(pretrained)	2e-4	--	20	0.0005
ResNet34(pretrained)	2e-4	--	20	0.0001
ResNet50(pretrained)	2e-4	--	20	0.0005
DenseNet121(pretrained)	2e-4	--	20	0.0005
Inception_v3(pretrained)	2e-4	--	30	0.0005

• *This is for test Model.

Model Result

Train Set

Model	Prec@1	Prec@5
AlexNet(pretrained)	93.4455	99.787
AlexNet(scratch)	99.3024	99.988
SketchANet(scratch)	86.3166	98.667
ResNet18(pretrained)	96.9899	99.954
ResNet34(pretrained)	97.1048	99.954
ResNet50(pretrained)	98.3049	99.988
DenseNet121(pretrained)	91.4301	99.596
Inception_v3(pretrained)	91.8802	99.706

Test Set

Model	Prec@1	Prec@5
Human	73.1	--
AlexNeti	68.6	--
AlexNetii	77.29	--
GoogLeNetii	80.85	--
AlexNet(pretrained)	70.850	90.050
AlexNet(scratch)	53.850	78.000
SketchANet(scratch)	68.700	88.900
ResNet18(pretrained)	77.800	94.650
ResNet34(pretrained)	79.100	95.050
ResNet50(pretrained)	78.300	95.300
DenseNet121(pretrained)	77.550	93.500
Inception_v3(pretrained)	76.550	93.750

• 1. Sketch-a-Net that Beats Humans
  2. The Sketchy Database: Learning to Retrieve Badly Drawn Bunnies

DPN, ShuffleNetG2, SENet18

Tools

• GetImageMean_Std

Get image dataset mean and standard deviation.

• SplitDataset

Split image dataset according to the train and val record txt file.

• ListAllImageName

Get all image name in dataset.