Sketch abstraction from "Goal-Driven Sequential Data Abstraction"

Introduction

This repo is official Tensorflow implementation of Goal-Driven Sequential Data Abstraction (ICCV 2019). It contains sketch abstraction part.

Repo structure

${SkechAbstraction}
|-- Source
|-- |-- InData
|   |-- Classifier
|   |   |-- Weights
|   |-- Agent
|   |   |-- Weights
|-- Output

• Download processed sketch dataset - 9 classes from QuickDraw [InData]
• Download pretrained classifier weights [Weights]
• Download pretrained agent weights [Weights]

Dependencies

The requirements.txt file lists all the required dependencies, which can be installed using:

pip install -r requirements.txt

Example usage

Train with CNN classifier and 25% budget size:

python main.py --trainFlag --classType 'CNN' --budgetVal 0.25 --outDir './Output/RNN-25/'

Test with CNN classifier and 25% budget size:

python main.py --classType 'CNN' --budgetVal 0.25 --agentLoading --agentLoadingFile './Source/Agent/Weights/Ag
ent_Weights_25_CNN.npy' --testStep 1

N.B. For training and testing of models with RNN classifier and budget size of 50%, please change the --classType and --budgetVal values accordingly.

Results

Category recognition (acc. %) of the abstracted sketches.

	RNN (25%)	RNN (50%)	CNN (25%)	CNN (50%)
Human	36.66	66.73	62.08	75.90
Random	22.67	45.65	41.06	65.47
DSA	38.36	67.89	65.05	81.50
DQSN	38.11	67.50	64.58	80.31
GDSA (Ours)	50.50	71.75	72.16*	86.15

Reference

@article{Muhammad_2019_ICCV_GDSA,
  author = {Umar Riaz Muhammad and Yongxin Yang and Timothy M. Hospedales and Tao Xiang and Yi-Zhe Song},
  title = {Goal-Driven Sequential Data Abstraction},
  booktitle = {The IEEE Conference on International Conference on Computer Vision (ICCV)},
  year = {2019}
}