SBIR-DASE-MEMS

This is the source code of Paper P. Lu, G. Huang, H. Lin, et.al. Domain-Aware SE Network for Sketch-based Image Retrieval with Multiplicative Euclidean Margin Softmax. In Proc. ACM Multimedia 2021. [paper]

This code is tested on Ubuntu 16.04 with python 3.7 and pytorch 1.3.1.

Framework

The framework of our model:

Datasets

• Download

  • TU-Berlin: Photos, Sketches
  • Sketchy: Photos, Sketches (We use sketches in folder tx_000100000000)

• File Structure

  - datasets
  	|- anno
  	|- data
  	|	|- TUBerlin
  	|	|	|- photos
  	|	|	|	|- airplane
  	|	|	|	|- ... (other categories)
  	|	|	|- sketches
  	|	|	|	|- airplane
  	|	|	|	|- ... (other categories)
  	|	|
  	|	|- sketchy
  	|	|	|- photos
  	|	|	|	|- airplane
  	|	|	|	|- ... (other categories)
  	|	|	|- sketches
  	|	|	|	|- airplane
  	|	|	|	|- ... (other categories)
  	|
  	|- Dataset.py
  	|- ...
  	
  

Demo Command

• Train ResNeXt-101 with MEMS loss

  # on TU-Berlin
  python main.py --save_path="tuberlin-resnext101-mems" --phase="train" --loss_type="mems" --margin=4
  # on Sketchy
  python main.py --save_path="sketchy-resnext101-mems" --phase="train" --dataset="sketchy" --loss_type="mems" --margin=4
  

• Using another loss

  python main.py --save_path="tuberlin-resnext101-cosface" --phase="train" --loss_type="cosface" --margin=0.35 --scale=32 --test_distance="angular"
  

• Using another network

  python main.py --save_path="tuberlin-alexnet-mems" --phase="train" --network="alexnet" --loss_type="mems" --margin=2 --batch_size=400
  

Pretrained Models

• tuberlin-resnext101-mems
• sketchy-resnext101-mems