Awesome

SegSwap

Pytorch implementation of paper "Learning Co-segmentation by Segment Swapping for Retrieval and Discovery"

Present in CVPR 2022 Image Matching Workshop and Transformers for Vision Workshop

[arXiv] [Project page] [Supplementary material] [Youtube Video][Slides]

If our project is helpful for your research, please consider citing :

@article{shen2021learning,
  title={Learning Co-segmentation by Segment Swapping for Retrieval and Discovery},
  author={Shen, Xi and Efros, Alexei A and Joulin, Armand and Aubry, Mathieu},
  journal={arXiv},
  year={2021}

Table of Content

• 0. Quickstart
• 1. Installation
• 2. Training Data Generation
• 3. Evaluation
  • 3.1 One-shot Art Detail Detection on Brueghel Dataset
  • 3.2 Place Recognition on Tokyo247 Dataset
  • 3.3 Place Recognition on Pitts30K Dataset
  • 3.4 Discovery on Internet Dataset
• 4. Train
• 5. Acknowledgement
• 6. ChangeLog
• 7. License

0. Quickstart

A quick start guide of how to use our code is available in demo/demo.ipynb

1. Installation

1.1. Dependencies

Our model can be learnt on a a single GPU Tesla-V100-16GB. The code has been tested in Pytorch 1.7.1 + cuda 10.2

Other dependencies can be installed via (tqdm, kornia, opencv-python, scipy) :

bash requirement.sh

1.2. Pre-trained MocoV2-resnet50 + cross-transformer (~300M)

Quick download :

cd model/pretrained
bash download_model.sh

2. Training Data Generation

2.1. Download COCO (~20G)

This command will download coco2017 training set + annotations (~20G).

cd data/COCO2017/download_coco.sh
bash download_coco.sh

2.2. Image Pairs with One Repeated Object

2.2.1 Generating 100k pairs (~18G)

This command will generate 100k image pairs with one repeated object.

cd data/
python generate_1obj.py --out-dir pairs_1obj_100k 

2.2.1 Examples of image pairs

2.2.2 Visualizing correspondences and masks of the generated pairs

This command will generate 10 pairs and visualize correspondences and masks of the pairs.

cd data/
bash vis_pair.sh

These pairs can be illustrated via vis10_1obj/vis.html

2.3. Image Pairs with Two Repeated Object

2.3.1 Generating 100k pairs (~18G)

This command will generate 100k image pairs with one repeated object.

cd data/
python generate_2obj.py --out-dir pairs_2obj_100k 

2.3.1 Examples of image pairs

2.3.2 Visualizing correspondences and masks of the generated pairs

This command will generate 10 pairs and visualize correspondences and masks of the pairs.

cd data/
bash vis_pair.sh

These pairs can be illustrated via vis10_2obj/vis.html

3. Evaluation

3.1 One-shot Art Detail Detection on Brueghel Dataset

3.1.1 Visual results: top-3 retrieved images

3.1.2 Data

Brueghel dataset has been uploaded in this repo

3.1.3 Quantitative results

The following command conduct evaluation on Brueghel with pre-trained cross-transformer:

cd evalBrueghel
python evalBrueghel.py --out-coarse out_brueghel.json --resume-pth ../model/hard_mining_neg5.pth --label-pth ../data/Brueghel/brueghelTest.json

Note that this command will save the features of Brueghel(~10G).

3.2 Place Recognition on Tokyo247 Dataset

3.2.1 Visual results: top-3 retrieved images

3.2.2 Data

Download Tokyo247 from its project page

Download the top-100 results used by patchVlad(~1G).

The data needs to be organised:

./SegSwap/data/Tokyo247
                    ├── query/
                        ├── 247query_subset_v2/
                    ├── database/
...

./SegSwap/evalTokyo
                    ├── top100_patchVlad.npy

3.2.3 Quantitative results

The following command conduct evaluation on Tokyo247 with pre-trained cross-transformer:

cd evalTokyo
python evalTokyo.py --qry-dir ../data/Tokyo247/query/247query_subset_v2 --db-dir ../data/Tokyo247/database --resume-pth ../model/hard_mining_neg5.pth

3.3 Place Recognition on Pitts30K Dataset

3.3.1 Visual results: top-3 retrieved images

3.3.2 Data

Download Pittsburgh dataset from its project page

Download the top-100 results used by patchVlad (~4G).

The data needs to be organised:

./SegSwap/data/Pitts
                ├── queries_real/
...

./SegSwap/evalPitts
                    ├── top100_patchVlad.npy

3.3.3 Quantitative results

The following command conduct evaluation on Pittsburgh30K with pre-trained cross-transformer:

cd evalPitts
python evalPitts.py --qry-dir ../data/Pitts/queries_real --db-dir ../data/Pitts --resume-pth ../model/hard_mining_neg5.pth

3.4 Discovery on Internet Dataset

3.4.1 Visual results

3.4.2 Data

Download Internet dataset from its project page

We provide a script to quickly download and preprocess the data (~400M):

cd data/Internet
bash download_int.sh

The data needs to be organised:

./SegSwap/data/Internet
                ├── Airplane100
                    ├── GroundTruth                
                ├── Horse100
                    ├── GroundTruth                
                ├── Car100
                    ├── GroundTruth                                

3.4.3 Quantitative results

The following commands conduct evaluation on Internet with pre-trained cross-transformer

cd evalInt
bash run_pair_480p.sh
bash run_best_only_cycle.sh

4. Training

Stage 1: standard training

Supposing that the generated pairs are saved in ./SegSwap/data/pairs_1obj_100k and ./SegSwap/data/pairs_2obj_100k.

Training command can be found in ./SegSwap/train/run.sh.

Note that this command should be able to be launched on a single GPU with 16G memory.

cd train
bash run.sh

Stage 2: hard mining

In train/run_hardmining.sh, replacing --resume-pth by the model trained in the 1st stage, than running:

cd train
bash run_hardmining.sh

5. Acknowledgement

We appreciate helps from :

• authors of Patch-NetVLAD who share their top-100 lists on Tokyo247 and Pitts30K with us.

• Dr. Relja Arandjelović for providing Tokyo247 and Pitts30K datasets.

• public code like Kornia

Part of code is borrowed from our previous projects: ArtMiner and Watermark

6. ChangeLog

• 27/10/22, add a demo
• 21/10/21, model, evaluation + training released
• 01/11/21, update arxiv link + supplmentary material

7. License

This code is distributed under an MIT LICENSE.

Note that our code depends on other libraries, including Kornia, Pytorch, and uses datasets which each have their own respective licenses that must also be followed.