wdr123/DARP-SBIR

code for ICDM submission

DARP-SBIR

Intro

This repository contains the source code implementation for ICDM submission paper Deep Reinforced Attention Regression for Partial Sketch Based Image Retrieval. The python files SBIR_*.py are the core framework codes to run this project under different settings, by integrating training code and evaluation code all in one file.

Dependencies

• numpy
• pickle
• pytorch>=1.5.0
• torchvision>=0.7.0
• tqdm
• tensorboard
• bresenham
• Pillow

Dataset

• Sketchy
• QMUL-Shoe-V2
• QMUL-Chair-V2

Pre-trained

• Pre-trained Embeddings

  • Train.pickle
  • Test.pickle

• Pre-trained Models

  • In folder ckpt

Results and Commands

To produce such results, run the corresponding entrance main files.

• Triplet Network: main_pure.py
• Triplet + Vanilla RL: main_RL.py
• Triplet + PPO: main_finetune.py
• Bootstrapped DQN: main_dqn.py
• DARP-SBIR: main_boot.py

For the other two datasets, use main_shoev2.py and main_sketchy.py