Matroid-LUPerson-extended

Introduction

This is a modified implementation of Unsupervised Pre-training for Person Re-identification (LUPerson) which is used for Appearance based Similarity search in Matroid.
The model is evaluated on three datasets, namely - Market1501, LaST and SYSU-30k. This documentation covers how to evaluate the model on all of these datasets as well as train the model on LaST and Market1501 dataset. It also covers steps to train and evaluate on a new dataset in the future.

Environment

Set up the conda environment

conda env create -f environment.yml
conda activate fastreid

Config file

The main parameters and arguments for training/evaluation are specified in config.yaml file such as

• Base model architecture
• Backbones
• input/output sizes
• Batch sizes
• Reranking
• Caching
• Parallelized evaluation

Checkout fast-reid/fastreid/config/defaults.py for comprehensive list of parameters set by config files.

Some important ones are:

• Model

MODEL:
  BACKBONE:
    WITH_IBN: False
    EXTRA_BN: True
  PIXEL_MEAN: [89.896, 79.200, 80.073]
  PIXEL_STD: [63.872, 64.305, 63.839]

• Dataset

DATASETS:
  NAMES: ("LaST",)
  TESTS: ("LaST",)
  KWARGS: 'normalize:False+train_mode:False+verbose:True'
  ROOT: "./datasets/last"

• Test options

TEST:
  EVAL_PERIOD: 60
  IMS_PER_BATCH: 128

If you want to enable re-ranking. Warning: slower and takes lot of CPU memory but more accurate results.

TEST:
  RERANK:
    ENABLED: True
    K1: 20
    K2: 6
    LAMBDA: 0.3

If you want to cache the generated features

TEST:
  CACHE:
    ENABLED: True
    CACHE_DIR: "logs/lup_moco/test/last/cache"

If you want to run eval on cached features with multiple worker threads on different cpus.

TEST:
  CACHE:
    REUSE_FEAT: True
    PARALLEL:
      ENABLED: True
      NUM_WORKERS: 12

Evaluation

Set variables

DATASET=market
PATH_TO_CHECKPOINT_FILE='<CHECKPOINT_PATH>'

Run eval script

python tools/train_net.py --eval-only --config-file <CONFIG_FILE_PATH> DATASETS.ROOT "datasets" DATASETS.KWARGS "data_name:${DATASET}" MODEL.WEIGHTS ${PATH_TO_CHECKPOINT_FILE} MODEL.DEVICE "cuda:0" OUTPUT_DIR "./logs/lup_moco/test/${DATASET}"

Specific settings for each dataset

• Market1501
  DATASET=market
  --config-file "./configs/CMDM/mgn_R50_moco.yml"

• LaST
  DATASET=last
  --config-file "./configs/LaST/mgn_R50_moco_cache_test.yml" to generate cached features
  --config-file "./configs/LaST/mgn_R50_moco_cache_reuse_parallelized_test.yml" to evaluate cached features

• SYSU-30k
  DATASET=last
  --config-file "./configs/SYSU-30k/mgn_R50_moco_cache_test.yml"

Training

Script to train on LaST dataset.

python tools/train_net.py --num-gpus 1 --config-file ./configs/LaST/mgn_R50_moco_train.yml SOLVER.CHECKPOINT_PERIOD 5

LUPerson (Original Paper documentation)

Unsupervised Pre-training for Person Re-identification (LUPerson).

The repository is for our CVPR2021 paper Unsupervised Pre-training for Person Re-identification.

LUPerson Dataset

LUPerson is currently the largest unlabeled dataset for Person Re-identification, which is used for Unsupervised Pre-training. LUPerson consists of 4M images of over 200K identities and covers a much diverse range of capturing environments.

Details can be found at ./LUP.

Pre-trained Models

Model	path
ResNet50	R50
ResNet101	R101
ResNet152	R152

Finetuned Results

For MGN with ResNet50:

Dataset	mAP	cmc1	path
MSMT17	66.06/79.93	85.08/87.63	MSMT
DukeMTMC	82.27/91.70	90.35/92.82	Duke
Market1501	91.12/96.16	96.26/97.12	Market
CUHK03-L	74.54/85.84	74.64/82.86	CUHK03

These numbers are a little different from those reported in our paper, and most are slightly better.

For MGN with ResNet101:

Dataset	mAP	cmc1	path
MSMT17	68.41/81.12	86.28/88.27	-
DukeMTMC	84.15/92.77	91.88/93.99	-
Market1501	91.86/96.21	96.56/97.03	-
CUHK03-L	75.98/86.73	75.86/84.07	-

The numbers are in the format of without RR/with RR.

Citation

If you find this code useful for your research, please cite our paper.

@article{fu2020unsupervised,
  title={Unsupervised Pre-training for Person Re-identification},
  author={Fu, Dengpan and Chen, Dongdong and Bao, Jianmin and Yang, Hao and Yuan, Lu and Zhang, Lei and Li, Houqiang and Chen, Dong},
  journal={Proceedings of the IEEE conference on computer vision and pattern recognition},
  year={2021}
}

News

We extend our LUPerson to LUPerson-NL with Noisy Labels which are generated from tracking algorithm, Please check for our CVPR22 paper Large-Scale Pre-training for Person Re-identification with Noisy Labels. And LUPerson-NL dataset is available at https://github.com/DengpanFu/LUPerson-NL