VFE_TPS

Code for our paper "Enhancing Visual Representation for Text-based Person Searching"

Usage

Requirements

PyTorch 2.0.0
torchvision 0.15.0
easydict
tqdm
prettytable

Prepare Datasets

The CUHK-PEDES dataset is proposed by the paper "Person Search with Natural Language Description" (Shuang Li, et al.) and can be downloaded from here. ICFG-PEDES dataset can be found here and RSTPReid dataset is here.

|-- your dataset root dir/
|   |-- <CUHK-PEDES>/
|       |-- imgs
|            |-- cam_a
|            |-- cam_b
|            |-- ...
|       |-- reid_raw.json
|
|   |-- <ICFG-PEDES>/
|       |-- imgs
|            |-- test
|            |-- train 
|       |-- ICFG_PEDES.json
|
|   |-- <RSTPReid>/
|       |-- imgs
|       |-- data_captions.json

Training

Directly run "run.sh" file

Testing

python test.py

Results

We use single Nvidia RTX 3090 GPU (24G) for training and testing

Results on CUHK-PEDES dataset

Method	Rank-1	Rank-5	Rank-10	mAP
AXM-Net	64.44	80.52	86.77	58.73
LGUR	64.21	81.94	87.93	-
IVT	65.59	83.11	89.21	-
CFine	69.57	85.93	91.15	-
TP-TPS	70.16	86.10	90.98	66.32
VGSG	71.38	86.75	91.86	-
VFE-TPS (ours)	72.47	88.24	93.24	64.26

Results on ICFG-PEDES dataset

Method	Rank-1	Rank-5	Rank-10	mAP
LGUR	59.02	75.32	81.56	-
MANet	59.44	76.80	82.75	-
CFine	60.83	76.55	82.42	-
TP-TPS	60.64	75.97	81.76	42.78
VFE-TPS (ours)	62.71	78.73	84.51	43.08

Results on RSTPReid dataset

Method	Rank-1	Rank-5	Rank-10	mAP
LBUL	45.55	68.20	77.85	-
IVT	46.70	70.00	78.80	-
CFine	50.55	72.50	81.60	-
TP-TPS	50.65	72.45	81.20	43.11
VFE-TPS (ours)	59.25	81.90	88.85	45.96

Acknowledgements

Our code is partially based on CLIP, IRRA. Sincerely appreciate for their contributions.