YizJia/DDPS - Githubissues

Introduction

This is the official implementation for "Disaggregation Distillation for Person Search" in TMM 2024.

We propose a novel compact and efficient person search model, namely Disaggregation Distillation for Person Search (DDPS), to distill the comprehensive knowledge from the teacher in a disaggregated manner. DDPS enhances the performance of the student by distilling from not only the task-oriented detection/re-id knowledge but also the discriminative intermediate features.
We propose three intuitive modules to achieve the objective of disaggregation distillation. On the one hand, DFD distinctively deals with person and background regions, enabling the student to capture critical knowledge during feature distillation. On the other hand, LD and CD contribute to the task-oriented distillation to boost the ultimate person search capabilities of the student.
Surprisingly, our compact person search models outperform the corresponding teacher models on both benchmarks.

DDPS

	Dataset	CUHK-SYSU	CUHK-SYSU	PRW	PRW
	Backbone	mAP	Top-1	mAP	Top-1
T	ResNet-50	91.96	92.72	41.43	79.24
S	ResNet-18	89.96	91.03	38.50	77.39
	w/ DDPS	92.57	93.45	41.93	79.58
	ResNet-18(0.5)	79.23	81.34	32.29	74.28
	w/DDPS	86.03	87.21	36.50	76.28
	MobileNetV2	89.60	91.00	39.85	79.97
	w/ DDPS	91.12	92.45	41.78	80.21

conda create -n ddps python=3.8 -y && conda activate ddps
pip install -r requirements.txt

Let's say $ROOT is the root directory.

Download CUHK-SYSU and PRW datasets, and unzip them to $ROOT/data
```
$ROOT/data
├── CUHK-SYSU
└── PRW
```
Following the link in the above table, download our pretrained model to anywhere you like, e.g., $ROOT/exp_cuhk
Run an inference demo by specifing the paths of checkpoint and corresponding configuration file. python demo.py --cfg $ROOT/exp_cuhk/config.yaml --ckpt $ROOT/exp_cuhk/epoch_19.pth You can checkout the result in demo_imgs directory.

bash job_train_base.sh

bash job_train_ddps.sh

Suppose the output directory is $ROOT/exp_cuhk. Test the trained model:

python train_ddps.py --cfg $ROOT/exp_cuhk/config.yaml --eval --ckpt $ROOT/exp_cuhk/epoch_19.pth

Thanks to the solid codebase from SeqNet.