TCNet

Implementation of Paper H. Lin, Y. Fu, P. Lu, et.al. TC-Net for iSBIR: Triplet Classification Network for instance-level Sketch Based Image Retrieval. In Proc. ACM Multimedia 2019.

Structure

The structure of our model.

Datasets

Datasets used in our paper. Your can modify the parameter in train.py to read different datasets.

QMUL Shoes / Chairs

• parameter obj

  --obj=shoes or --obj=chairs

Shoes v2

• file structure

  - QUML_v2
  	|- ShoeV2_photo
  	|	|- 1135045020.png
  	|	|- ...
  	|
  	|- ShoeV2_sketch
  	|	|- 1135045020_1.png
  	|	|- ...
  	|
  	|- photo_train.txt
  	|- photo_test.txt
  	|- sketch_train.txt
  	|- sketch_test.txt
  

• parameter obj,data_root

  --obj=shoes_v2 --data_root=/home/xxx/dataset/sketch/sbir_qian/QUML_v2

Sketchy

• file structure

  - sketchy
  	|- 256x256
  	|	|- photo
  	|	|	|- tx_000100000000
  	|	|	|	|- airplane
  	|	|	|	|	|- n02691156_507.jpg
  	|	|	|	|	|- ... (othre photos)
  	|	|	|	|- ... (other categories)
  	|	|	|- tx_(does not matter)
  	|	|	
  	|	|- sketch
  	|	|	|- tx_000100000000
  	|	|	|	|- airplane
  	|	|	|	|	|- n02691156_507-1.jpg
  	|	|	|	|	|- ... (other sketches)
  	|	|	|	|- ... (other categories)
  	|	|	|- tx_(does not matter)
  	|
  	|- info
  	|	|- invalid-ambiguous.txt
  	|	|- invalid-context.txt
  	|	|- invalid-error.txt
  	|	|- invalid-pose.txt
  	|	|- testset.txt
  

• parameter obj,data_root

  --obj=sketchy --data_root=/home/xxx/dataset/sketch/sketchy

Demo command

python train.py --obj=shoes_v2 --data_root=/home/lp_user/dataset/sketch/sbir_qian/QUML_v2 --model_type=densenet --feat_dim=1024 --loss_type='triplet,centre' --loss_ratio='0.13,0.0013' --flag=shoesv2-tpl_ctr-densebn

• training from pretrained model

  set --phase=train_continue and make sure parameters flag and obj are identical with the previous model.