Sketch Triplet Networks.

A PyTorch Implementation for Sketch Triplet Networks.

Model Configuration

• Optimizer
  • Adam

DataSet

• shoes
• chairs

Model Parameters

Model branch	pretrained	Loss Function	lr	clip_grad_norm(max_norm)	learning rate decay	weight_decay	Margin
SketchANet	TU-Berlin	Triplet Loss	2e-5	--	20	0.0005(shoes) 0.0005-0.001(chairs)	0.3
AlexNet	T(ImageNet)		2e-4	1.0	100	0.0003	0.3
ResNet18	T	MarginRankingLoss	2e-6	10.0	20	0.05	0.3
ResNet18	T(TU-Berlin)	TripletMarginLoss	2e-6	10.0	20	0.01	0.3

Model Result

Shoes

Train Set(ranking)

Model branch	pretrained	Loss Function	prec	mprec
AlexNet	T	TripletMarginLoss
ResNet18	T	MarginRankingLoss
ResNet18	T(ImageNet)	TripletMarginLoss
ResNet18	T(TU-Berlin)	TripletMarginLoss

Test Set(ranking)

Model branch	pretrained	Loss Function	prec	mprec
SketchANet	TU-Berlin	Triplet Loss
AlexNet	T	TripletMarginLoss	61.76	15.34
ResNet18	T	MarginRankingLoss
ResNet18	T(ImageNet)	TripletMarginLoss
ResNet18	T(TU-Berlin)	TripletMarginLoss

Test Set(retrieval)

Model branch	pretrained	Loss Function	Rank@1	Rank@5	Rank@10	corr
Origini		Triplet Loss	39.13	--	87.83	69.49
Originii	ImageNet(edge)+TU-Berlin	Triplet Loss(square_distance)	52.174	--	92.174	--
SketchANet	TU-Berlin	Triplet Loss	45.217	77.391	82.609	72.15
AlexNet	ImageNet+TU-Berlin	Triplet Loss	45.217	74.783	86.087	73.70
AlexNet	T	TripletMarginLoss
ResNet18	TU-Berlin	TripletLoss	26.957	51.304	64.348	64.54
ResNet18	T	MarginRankingLoss
ResNet18	TU-Berlin	MarginRankingLoss	29.565	50.435	69.565	64.21
ResNet18	ImageNet	TripletMarginLoss
ResNet18	TU-Berlin	TripletMarginLoss	25.217	53.043	65.217	64.79
ResNet18	TU-Berlin	TripletMarginLoss + embedded_norm

Chairs

Train Set(ranking)

Test Set(ranking)

Model branch	pretrained	Loss Function	prec	mprec
SketchANet	TU-Berlin	Triplet Loss	74.46	51.09

Test Set(retrieval)

Model branch	pretrained	Loss Function	Rank@1	Rank@5	Rank@10	corr
Origini		Triplet Loss	69.07	--	97.94	72.30
Originii	ImageNet(edge)+TU-Berlin	Triplet Loss(square_distance)	72.16	--	98.96	--
SketchANet	TU-Berlin	Triplet Loss	76.289	91.753	92.784	73.45
AlexNet	ImageNet+TU-Berlin	Triplet Loss	63.918	87.629	92.784	73.13
ResNet18	ImageNet+TU-Berlin	Triplet Loss	61.856	87.629	93.814	76.01

[Reference]

• 1. Sketch Me That Shoe
  2. Deep Spatial-Semantic Attention for Fine-Grained Sketch-Based Image Retrieval