PTL: Progressive Transfer Learning for Person Re-identification

Introduction

PTL is a model fine-tuning method for deep neural networks. It provides an efficient solution for the model fine-tuning task, and can improve the performance of the pre-trained model on the target dataset significantly.

This project is the implementation of the Batch-related Convolutional Cell (BConv-Cell) and the MGN_PTL network of our IJCAI-2019 paper - Progressive Transfer Learning for Person Re-identification.

PTL method has been used by the CityBrain Group (Damo Academy, Alibaba Group) to help improve the model performance when using the pre-trained ReID models in a newly emerged scenario.

BConv-Cell

The key component of the MGN_PTL network is the BConv-Cell in bconv_cell.py
The BConv-Cell can integrate with most deep neural networks to improve the model performance when using mini-batch training.
This project only provides an example of its usage, feel free to explore.

Performance

Datasets

Market-1501

Download using:

wget http://188.138.127.15:81/Datasets/Market-1501-v15.09.15.zip <path/to/where/you/want>
unzip <path/to/>/Market-1501-v15.09.15.zip

CUHK03
1. Download cuhk03 dataset from here
2. Unzip the file and you will get the cuhk03_release dir which include cuhk-03.mat
3. Download "cuhk03_new_protocol_config_detected.mat" from here and put it with cuhk-03.mat. We need this new protocol to split the dataset.
```
python utils/transform_cuhk03.py --src <path/to/cuhk03_release> --dst <path/to/save>
```
NOTICE: You need to change num_classes in network depend on how many people in your train dataset! e.g. 751 in Market1501.

The data structure should look like:

data/
    bounding_box_train/
    bounding_box_test/
    query/

Compared person ReID methods

DML: Deep Mutual Learning
HA-CNN: Harmonious Attention Network for Person Re-identification
PCB: Beyond Part Models: Person Retrieval with Refined Part Pooling (and a Strong Convolutional Baseline)
PCB+RPP: Beyond Part Models: Person Retrieval with Refined Part Pooling (and a Strong Convolutional Baseline)
MGN: Learning Discriminative Features with Multiple Granularities for Person Re-Identification

Results

Method	Market-1501		DukeMTMC-reID		CUHK03(Detected)		CUHK03(Labelled)
	mAP	CMC-1	mAP	CMC-1	mAP	CMC-1	mAP	CMC-1
DML	70.51	89.34	-	-	-	-	-	-
HA-CNN	75.70	91.20	63.80	80.50	38.60	41.70	41.00	44.40
PCB	77.30	92.40	65.30	81.90	54.20	61.30	-	-
PCB+RPP	81.60	93.80	69.20	83.30	57.50	63.70	-	-
MGN	86.90	95.70	78.40	88.70	66.00	66.80	67.40	68.00
MGN(reproduced)	85.80	94.60	77.07	87.70	69.41	71.64	72.96	74.07
MGN_PTL	87.34	94.83	79.16	89.36	74.22	76.14	77.31	79.79

NOTICE: The MGN(reproduced) is the reproduction of MGN. To our best knowledge, the official implementation of MGN has not released yet. Hence, the MGN_PTL network used the MGN(reproduced) as backbone network. The code for MGN(reproduced) is in mgn.py

RUN

Prerequisites

cudnn 7
CUDA 9
Pytorch v0.4.1
Python 2.7
torchvision
scipy
numpy
scikit_learn

GPU usage

We used one Tesla P100 GPU in our experiments

To run the MGN with batchid=4 and batchimage=4 cost 7819 MiB
To run the MGN_PTL with batchid=4 and batchimage=4 cost 8819 MiB

Weights

Pretrained weight download from TODO (Currently unavailable, the weight file will be released later)

Train

You can specify more parameters in opt.py

Train MGN_PTL

CUDA_VISIBLE_DEVICES=0 python train_eval.py --arch mgn_ptl --mode train --usegpu --project_name 'temp_project' --data_path <path/to/Market-1501-v15.09.15> --lr 2e-4 --batchid 4 --epoch 450

Train MGN

CUDA_VISIBLE_DEVICES=0 python train_eval.py --arch mgn --mode train --usegpu --project_name 'temp_project' --data_path <path/to/Market-1501-v15.09.15> --lr 2e-4 --batchid 4 --epoch 450

Evaluate

Use pretrained weight or your trained weight

Evaluate MGN_PTL

CUDA_VISIBLE_DEVICES=0 python train_eval.py --arch mgn_ptl --mode evaluate --usegpu --weight <path/to/weight/weight_name.pt> --data_path <path/to/Market-1501-v15.09.15>

Evaluate MGN

CUDA_VISIBLE_DEVICES=0 python train_eval.py --arch mgn --mode evaluate --usegpu --weight <path/to/weight/weight_name.pt> --data_path <path/to/Market-1501-v15.09.15>

Reference

Reference to cite when you use PTL in a research paper:

@inproceedings{ijcai2019-586,
  title     = {Progressive Transfer Learning for Person Re-identification},
  author    = {Yu, Zhengxu and Jin, Zhongming and Wei, Long and Guo, Jishun and Huang, Jianqiang and Cai, Deng and He, Xiaofei and Hua, Xian-Sheng},
  booktitle = {Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, {IJCAI-19}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},             
  pages     = {4220--4226},
  year      = {2019},
  month     = {7},
  doi       = {10.24963/ijcai.2019/586},
  url       = {https://doi.org/10.24963/ijcai.2019/586},
  }

License

PTL is MIT-licensed.

aliang-cn / ptl Goto Github PK

ptl's Introduction