sillymistake / Partial-Person-ReID

Since the testing of the previous code was slow, we optimized it to speed up the model testing.

Partial-Person-ReID

---

The source code: Spatial Feature Reconstruction with Pyramid Pooling for Partial Person Re-identification CVPR18: Deep Spatial Feature Reconstruction for Partial Person Re-identification: Alignment-free Approach, Arxiv18

The project provides the training and testing code for partial person re-id, using Pytorch

Instllation

---

It's recommended that you create and enter a python virtual environment, if versions of the packages required here conflict with yours.

Other packages are specified in requirements.txt

Daset Preparation

Inspired by Houjing Huang's person-reid-triplet-loss-baseline project, you can follow his guidance.

Experiment Setting:

1. Backbone: ResNet-50, stride = 1 in the last conv block.
2. Input image size: 384 & times 192

Person Re-identification

Result on Market1501

python script/experiment/train.py \
--dataset market1501 \
--partial_dataset others\
--Spatial_train False \
--total_epochs 400 

Method	Rank-1 (Single query)	mAP	Rank-1 (Multi query)	mAP
Baseline	88.18	73.85	92.25	80.96
SFR	93.04	81.02	94.84	85.47

Result on CUHK03

python script/experiment/train.py \
--dataset cuhk03 \
--partial_dataset others\
--Spatial_train False \
--total_epochs 400 

Method	Rank-1 (Labeled)	mAP	Rank-1 (Detected)	mAP
Baseline	62.14	58.47	60.43	54.24
SFR	67.29	61.47	63.86	58.97

Result on Duke

python script/experiment/train.py \
--dataset duke \
--partial_dataset others\
--Spatial_train False \
--total_epochs 400 

Method	Rank-1 (Labeled)	mAP
Baseline	80.48	64.80
SFR	84.83	71.24

Partial Person Re-identification

The link of Partial REID and Partial iLIDS datasets: Baidu Cloud.

Before run the code, you should revise the path in Partial_REID_test.py and Partial_iLIDS_test.py to your path.

Result on Partial REID

python script/experiment/train.py \
--dataset market1501 \
--partial_dataset Partial_REID\
--Spatial_train False \
--total_epochs 400 

Method	Rank-1	Rank-5
Baseline	54.80	80.20
SFR	66.20	86.67

Result on Partial iLIDS

python script/experiment/train.py \
--dataset market1501 \
--partial_dataset Partial_iLIDS\
--Spatial_train False \
--total_epochs 400 

Method	Rank-1	Rank-5
Baseline	46.22	74.79
SFR	63.87	86.55

if you want to add the spatial feature reconstruction (SFR) in training term, please set Spatial_train=True, but it would increase the training time.

Citing Spatial Feature Reconstruction

If you find SFR is useful in your research, pls consider citing:

@InProceedings{He_2018_CVPR,
author = {He, Lingxiao and Liang, Jian and Li, Haiqing and Sun, Zhenan}, 
title = {Deep Spatial Feature Reconstruction for Partial Person Re-Identification: Alignment-Free Approach},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2018}
} ```