• AGW Baseline: Deep Learning for Person Re-identification: A Survey and Outlook
• Contents
• • Model Description
  • Model Architecture
  • Dataset
  • • Market1501
  • Environment Requirements
  • Quick Start
  • • Running scripts
  • Script Description
  • • Script and Sample Code
    • Script Parameters
    • • Training Script Parameters
  • Training Process
  • • on Ascend
    • on GPU
  • Evaluation Process
  • • on Ascend
    • on GPU
  • Inference Process
  • • Export MindIR
    • result
  • Model Description
  • • Performance
    • • Market1501 Training Performance
      • Market1501 Evaluation Performance
  • Description of Random Situation
  • ModelZoo Homepage

AGW Baseline: Deep Learning for Person Re-identification: A Survey and Outlook

Contents

• AGW Baseline: Deep Learning for Person Re-identification: A Survey and Outlook
• Contents
  • Model Description
  • Model Architecture
  • Dataset
    • Market1501
  • Environment Requirements
  • Quick Start
    • Running scripts
  • Script Description
    • Script and Sample Code
    • Script Parameters
      • Training Script Parameters
  • Training Process
  • Evaluation Process
  • Inference Process
    • Export MindIR
    • result
  • Model Description
    • Performance
      • Market1501 Training Performance
      • Market1501 Evaluation Performance
  • Description of Random Situation
  • ModelZoo Homepage

Model Description

ReID Strong Baseline proposes a novel neck structure named as batch normalization neck (BNNeck).
BNNeck adds a batch normalization layer after global pooling layer to separate metric and classification
losses into two different feature spaces because we observe they are inconsistent in one embedding space.
Extended experiments show that BNNeck can boost the baseline.

Paper: Luo H., Jiang W., et al. “A Strong Baseline and Batch Normalization Neck for Deep Person Re-identification”. CVPRW2019, Oral.

Model Architecture

Model uses ResNet50 as backbone. BNNeck adds a BN layer after features and before classifier FC layers.
The BN and FC layers are initialized through Kaiming initialization.

Dataset

Market1501

Market1501 dataset is used to train and test model.
Market1501 contains 32,668 images of 1,501 labeled persons of six camera views.
There are 751 identities in the training set and 750 identities in the testing set.
In the original study on this proposed dataset, the author also uses mAP as the evaluation criteria to test the algorithms.

Data structure:

Market-1501-v15.09.15  
├── bounding_box_test [19733 entries]
├── bounding_box_train [12937 entries]
├── gt_bbox [25260 entries]
├── gt_query [6736 entries]
├── query [3369 entries]
└── readme.txt

Environment Requirements

• Hardware（GPU）
  • Prepare hardware environment with GPU processor.
• Framework
  • MindSpore
• For more information, please check the resources below：
  • MindSpore Tutorials
  • MindSpore Python API

Quick Start

Running scripts

Model uses pre-trained backbone ResNet50 trained on ImageNet2012. Link

After dataset preparation, you can start training and evaluation as follows:

(Note that you must specify dataset path in configs/market1501_config.yml)

# run training example
bash scripts/run_standalone_train_gpu.sh  ./configs/market1501_config.yml 0 /path/to/dataset/ /path/to/output/ /path/to/pretrained_resnet50.pth

# run evaluation example
bash scripts/run_eval_gpu.sh ./configs/market1501_config.yml /path/checkpoint_file /path/to/dataset/

Script Description

Script and Sample Code

AGW
├── README.md
├── configs
│   ├── dukemtmc_config.yml
│   └── market1501_config.yml
├── eval.py
├── export.py
├── requirements.txt
├── scripts
│   ├── run_distribute_train_ascend.sh
│   ├── run_eval_ascend.sh
│   ├── run_eval_gpu.sh
│   ├── run_standalone_train_ascend.sh
│   └── run_standalone_train_gpu.sh
├── src
│   ├── __init__.py
│   ├── callbacks.py
│   ├── center_loss.py
│   ├── dataset.py
│   ├── datasets
│   │   ├── __init__.py
│   │   ├── bases.py
│   │   ├── dukemtmcreid.py
│   │   └── market1501.py
│   ├── loss.py
│   ├── lr_schedule.py
│   ├── metric_utils.py
│   ├── model
│   │   ├── __init__.py
│   │   ├── agw.py
│   │   ├── cell_wrapper.py
│   │   ├── resnet_nl.py
│   │   └── vit.py
│   ├── model_utils
│   │   ├── __init__.py
│   │   ├── config.py
│   │   ├── device_adapter.py
│   │   ├── local_adapter.py
│   │   └── moxing_adapter.py
│   ├── sampler.py
│   └── triplet_loss.py
├── train.py
└── train_ascend_qizhi.p

Script Parameters

Training Script Parameters

usage: train.py  --config_path CONFIG_PATH  [--device_target DEVICE]

Training Process

• Set options in configs/market1501_config.yaml or configs/dukemtmc_config.yaml,
  including paths, learning rate and network hyperparameters.

on Ascend

• Run run_distribute_train_Ascend.sh for distributed training of ReID Strong Baseline model.

  The RANK_TABLE_FILE is placed under scripts/

  bash scripts/run_distribute_train_Ascend.sh CONFIG_PATH DATA_DIR OUTPUT_PATH PRETRAINED_RESNET50 RANK_TABLE_FILE RANK_SIZE
  

on GPU

• Run run_standalone_train_gpu.sh for non-distributed training of model.

  bash scripts/run_standalone_train_gpu.sh CONFIG_PATH DEVICE_ID DATA_DIR OUTPUT_PATH PRETRAINED_RESNET50
  

Evaluation Process

• Set options in market1501_config.yaml.

on Ascend

• Run bash scripts/run_eval_Ascend.sh for evaluation of ReID Strong Baseline model.

  bash scripts/run_eval_Ascend.sh CONFIG_PATH CKPT_PATH DATA_DIR
  

on GPU

• Run bash scripts/run_eval_gpu.sh for evaluation of ReID Strong Baseline model.

  bash scripts/run_eval_gpu.sh CONFIG_PATH CKPT_PATH DATA_DIR
  

Inference Process

Export MindIR

python export.py --config_path [CONFIG_PATH] --ckpt_file [CKPT_PATH] --file_name [FILE_NAME] --file_format [FILE_FORMAT]

options:
    --config_path              path to .yml config file
    --ckpt_file                checkpoint file
    --file_name                output file name
    --file_format              output file format, choices in ['MINDIR']

The ckpt_file and config_path parameters are required,
FILE_FORMAT should be in "MINDIR"

result

Inference result will be shown in the terminal

Model Description

Performance

Market1501 Training Performance

Parameters	Ascend
Resource	1x Ascend 910
uploaded Date	2023/04/23
MindSpore Version	1.8.1
Dataset	Market-1501
Training Parameters	max_epoch=120, ids_per_batch=16, ims_per_id=4, metric='euclidean'
Optimizer	Adam, SGD
Loss Function	Smooth Identity, WRT, Center
Speed	444ms/step
Loss	0.23
Checkpoint for inference	400.58MB (.ckpt file)
Scripts	scripts

Parameters	Ascend
Resource	4x Ascend 910
uploaded Date	2023/04/23
MindSpore Version	1.8.1
Dataset	Market-1501
Training Parameters	max_epoch=120, ids_per_batch=16, ims_per_id=4, metric='euclidean', distributed=1
Optimizer	Adam, SGD
Loss Function	Smooth Identity, WRT, Center
Speed	444ms/step
Loss	0.23
Checkpoint for inference	400.52MB (.ckpt file)
Scripts	scripts

Market1501 Evaluation Performance

Parameters	Ascend
Resource	1x Ascend 910
Uploaded Date	2023/04/24
MindSpore Version	1.8.1
Dataset	Market-1501
batch_size	32
outputs	mAP, Rank-1, mINP
Accuracy	mAP: 0.8912 mINP:0.6873 rank1: 0.9516

Parameters	Ascend
Resource	4x Ascend 910
Uploaded Date	2023/04/24
MindSpore Version	1.8.1
Dataset	Market-1501
batch_size	32
outputs	mAP, Rank-1, mINP
Accuracy	mAP: 0.8673 mINP:0.6336 rank1: 0.9448

Description of Random Situation

There are six random situations:

• Shuffle of the persons in the dataset.
• Select random images for persons.
• Random flip images.
• Random crop images.
• Random erasing images.
• Initialization of some model weights.

Some seeds have already been set in train.py and sampler.py to avoid the randomness of dataset shuffle and weight initialization.

ModelZoo Homepage

Please check the official homepage.