Home

Awesome

Zero-Shot Everything Sketch-Based Image Retrieval, and in Explainable Style

Python 3.6 Pytorch 1.10 MIT License

In this repository, you can find the official PyTorch implementation of Zero-Shot Everything Sketch-Based Image Retrieval, and in Explainable Style, CVPR2023, Highlight.

Authors: Fengyin Lin, Mingkang Li, Da Li, Timothy Hospedales, Yi-Zhe Song and Yonggang Qi, Beijing University of Posts and Telecommunications, Samsung AI Centre Cambridge, University of Edinburgh, SketchX CVSSP University of Surrey.

Abstract: This paper studies the problem of zero-short sketch-based image retrieval (ZS-SBIR), however with two significant differentiators to prior art (i) we tackle all variants (inter-category, intra-category, and cross datasets) of ZS-SBIR with just one network (“everything”), and (ii) we would really like to understand how this sketch-photo matching operates (“explainable”). Our key innovation lies with the realization that such a cross-modal matching problem could be reduced to comparisons of groups of key local patches – akin to the seasoned “bag-of-words” paradigm. Just with this change, we are able to achieve both of the aforementioned goals, with the added benefit of no longer requiring external semantic knowledge. Technically, ours is a transformer-based cross-modal network, with three novel components (i) a self-attention module with a learnable tokenizer to produce visual tokens that correspond to the most informative local regions, (ii) a cross-attention module to compute local correspondences between the visual tokens across two modalities, and finally (iii) a kernel-based relation network to assemble local putative matches and produce an overall similarity metric for a sketch-photo pair. Experiments show ours indeed delivers superior performances across all ZS-SBIR settings. The all important explainable goal is elegantly achieved by visualizing cross-modal token correspondences, and for the first time, via sketch to photo synthesis by universal replacement of all matched photo patches.

Fig.1

Datasets

Please download SBIR datasets from the official websites or Google Drive and tar -zxvf dataset to the corresponding directory in ./datasets. We provide train and test splits for different datasets.

Sketchy

Please go to the Sketchy official website, or download the dataset from Google Drive.

TU-Berlin

Please go to the TU-Berlin official website, or download the dataset from Google Drive.

QuickDraw

Please go to the QuickDraw official website, or download the dataset from Google Drive.

Installation

The requirements of this repo can be found in requirements.txt.

conda create -n zse-sbir python=3.6
pip install -r requirements.txt
# or the same pytorch versioin as ours
pip install torch==1.10.0+cu113 torchvision==0.11.1+cu113 torchaudio==0.10.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html

Train

Pretrained ViT backbone

The pre-trained ViT model on ImageNet-1K is provided on Google Drive. You should place sam_ViT-B_16.pth in ./model and modify line 190 in ./model/sa.py to absolute path if necessary.

Haperparameters

Here is a list of full options for the model:

# dataset
data_path,            # path to load datasets.
dataset,              # choose a dataset for train or eval.
test_class,           # choose a zero-shot split of dataset.

# model
cls_number,           # class number if necessary, 100 as default.
d_model,              # feature dimension, 768 as default.
d_ff,                 # fead-forward layer dimension, 1024 as default.
head,                 # number of ca encoder head, 8 as default.
number,               # number of ca encoder layer, 1 as default.
pretrained,           # whether to use pretrained ViT model, true as default.
anchor_number,        # number of anchor in rn network, 49 as default.

# train
save, -s,             # path to save checkpoints.
batch,                # batch size, 15 as default.
epoch,                # train epoch, 30 as default.
datasetLen,           # data pair for train per epoch, 10000 as default.
learning_rate,        # learning rate, 1e-5 as default.
weight_decay,         # weight decay, 1e-2 as default.

# test
load, -l,             # path to load checkpoints.
retrieval, -r,        # test method, rn for ret-token and sa for cls-token, use rn as default.
testall,              # whether use all test data, suggesting false for train, true for test.
test_sk,              # number of sketches per loop during test, 20 as default.
test_im,              # number of images per loop during test, 20 as default.
num_workers,          # dataloader num workers, 4 as default.

# other
choose_cuda, -c,      # cuda to use, 0 as default.
seed,                 # random seed, 2021 as default.

Train ZSE-SBIR

Here is a quick start for training the network on Sketchy Ext. Please pay attention to modifying data path and save path before run.

python -u train.py 
# or use nohup command
nohup python -u train.py > sketchy_ext.log 2>&1 &

Train model on Sketchy Ext.

python -u train.py --data_path [./datasets] \
                   --dataset sketchy_extend \ 
                   --test_class test_class_sketchy25 \ 
                   --batch 15 \ 
                   --epoch 30 \ 
                   -s [./checkpoints/sketchy_ext] \
                   -c 0 \ 
                   -r rn

Train model on TU-Berlin Ext.

python -u train.py --data_path [./datasets] \
                   --dataset tu_berlin \ 
                   --test_class test_class_tuberlin30 \ 
                   --batch 15 \ 
                   --epoch 30 \ 
                   -s [./checkpoints/tuberlin_ext] \
                   -c 0 \ 
                   -r rn \

Train model on QuickDraw Ext.

python -u train.py --data_path [./datasets] \
                   --dataset Quickdraw \ 
                   --test_class Quickdraw \ 
                   --batch 15 \ 
                   --epoch 30 \ 
                   -s [./checkpoints/quickdraw_ext] \
                   -c 0 \ 
                   -r rn

Evaluation

Our Trained Model

The trained model on Sketchy Ext is provided on Google Drive. You should place best_checkpoint.pth in ./checkpoint/sketchy_ext and modify load path --load for example.

Evaluate ZSE-SBIR

Here is a quick start for evaluating the network on Sketchy Ext. Please pay attention to modifying data path and save path before run.

# use rn-score for zs-sbir, and use all test data.
python -u test.py -r rn -- testall
# use ret-token for zs-sbir, which is quite faster.
python -u test.py -r sa -- testall
# or use nohup command
nohup python -u test.py -r rn --testall > test_sketchy_ext.log 2>&1 &

Evaluate model on Sketchy Ext.

python -u test.py --data_path [./datasets] \
                  --dataset sketchy_extend \
                  --test_class test_class_sketchy25 \ 
                  -l [./checkpoints/sketchy_ext/best_checkpoint.pth] \
                  -c 0 \ 
                  -r rn \ 
                  --testall

Evaluate model on TU-Berlin Ext.

python -u test.py --data_path [./datasets] \
                  --dataset tu_berlin \
                  --test_class test_class_tuberlin30 \ 
                  -l [./checkpoints/tuberlin_ext/best_checkpoint.pth] \
                  -c 0 \ 
                  -r rn \ 
                  --testall

Evaluate model on QuickDraw Ext.

python -u test.py --data_path [./datasets] \
                  --dataset Quickdraw \
                  --test_class Quickdraw \ 
                  -l [./checkpoints/quickdraw_ext/best_checkpoint.pth] \
                  -c 0 \ 
                  -r rn \ 
                  --testall

License

This project is released under the MIT License.

Citation

If you find this repository useful for your research, please use the following.

@inproceedings{zse-sbir-cvpr2023,
    title={Zero-Shot Everything Sketch-Based Image Retrieval, and in Explainable Style},
    author={Fengyin Lin, Mingkang Li, Da Li, Timothy Hospedales, Yi-Zhe Song and Yonggang Qi},
    booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2023}
}