On the Unreasonable Effectiveness of Centroids in Image Retrieval

Official code repository for paper "On the Unreasonable Effectiveness of Centroids in Image Retrieval".
Paper accepted to ICONIP 2021 conference.

Get Started

The whole model is implemented in PyTorch-Lightning framework.

cd to a directory where you want to clone this repo
Run git clone https://github.com/mikwieczorek/centroids-reid.git
Install required packages pip install -r requirements.txt
Download pre-trained weights into models/ directory for:
- Resnet50 from here: [link]
- Resnet50-IBN-A from here: [link]
Prepare datasets:

Market1501
- Extract dataset and rename to market1501 inside /data/
- The data structure should be following:
```
/data
    market1501
        bounding_box_test/
        bounding_box_train/
        ......
```
DukeMTMC-reID
- Extract dataset to /data/ directory
- The data structure should be following:
```
/data
    DukeMTMC-reID
       	bounding_box_test/
       	bounding_box_train/
       	......
```
Street2Shop & Deep Fashion (Consumer-to-shop)
1. These fashion datasets require the annotation data in COCO-format with additional fields in annotations
```
JSON:{
    'images' : [...],
    'annotations': [
                {...,
                'pair_id': 100,         # an int type
                'source': 'user'        # 'user' or 'shop'
                },
                ...
            ]
}
```
2. The product images should be pre-cropped to the given input format (either 256x128 or 320x320) using original images and provided bounding boxes to allow faster training.
Path to the data root and JSON files (only for Street2shop and Deep Fashion) can be adjusted by passing the paths as parameters to train scripts

You can familiarize yourself with the detailed configuration and its meaning in config.defaults.py, which includes all parameters available to the user.

Train

Each Dataset and Model has its own train script.
All train scripts are in train_scirpts folder with corresponding dataset name.

Example run command to train CTL-Model on DukeMTMC-reID

CUDA_VISIBLE_DEVICES=3 ./train_scripts/dukemtmc/train_ctl_model_s_r50_dukemtmc.sh

CUDA_VISIBLE_DEVICES controls which GPUs are visible to the scripts.

By default all train scripts will launch 3 experiments.

Test

To test the trained model you can use provided scripts in train_scripts, just two parameters need to be added:

TEST.ONLY_TEST True \  
MODEL.PRETRAIN_PATH "path/to/pretrained/model/checkpoint.pth"

Example train script for testing trained CTL-Model on Market1501

python train_ctl_model.py \
--config_file="configs/256_resnet50.yml" \
GPU_IDS [0] \
DATASETS.NAMES 'market1501' \
DATASETS.ROOT_DIR '/data/' \
SOLVER.IMS_PER_BATCH 16 \
TEST.IMS_PER_BATCH 128 \
SOLVER.BASE_LR 0.00035 \
OUTPUT_DIR './logs/market1501/256_resnet50/' \
SOLVER.EVAL_PERIOD 40 \
TEST.ONLY_TEST True \
MODEL.PRETRAIN_PATH "logs/market1501/256_resnet50/train_ctl_model/version_0/checkpoints/epoch=119.ckpt"

Trained model weights

Weights of trained CTL-Model on Market1501 and DuketMTMC-Reid are avaialable here [link]

Inference

To run inference on an image dataset with a trained model you can use provided scripts in inference folder. More info can be found in inference/README.md.

Fashion dataset to COCO-ReID format preparation

Fashion datasets require transforming to COCO-Reid format. In scripts folder there is code necessary to prepare raw Street2Shop and DeepFashion annotation and images to correct format. More info can be found in scripts/README.md.

Citation

Wieczorek M., Rychalska B., Dąbrowski J. 
(2021) On the Unreasonable Effectiveness of Centroids in Image Retrieval.
In: Mantoro T., Lee M., Ayu M.A., Wong K.W., Hidayanto A.N. (eds) 
Neural Information Processing. ICONIP 2021.
Lecture Notes in Computer Science, vol 13111. Springer, Cham. https://doi.org/10.1007/978-3-030-92273-3_18

@article{Wieczorek2021OnTU,
  title={On the Unreasonable Effectiveness of Centroids in Image Retrieval},
  author={Mikolaj Wieczorek and Barbara Rychalska and Jacek Dabrowski},
  journal={ArXiv},
  year={2021},
  volume={abs/2104.13643}
}

mikwieczorek/centroids-reid