Efficient Crowd Counting via Structured Knowledge Transfer (ACM MM 2020)

Crowd counting is an application-oriented task and its inference efficiency is crucial for real-world applications. However, most previous works relied on heavy backbone networks and required prohibitive run-time consumption, which would seriously restrict their deployment scopes and cause poor scalability. To liberate these crowd counting models, we propose a novel Structured Knowledge Transfer (SKT) framework, which fully exploits the structured knowledge of a well-trained teacher network to generate a lightweight but still highly effective student network.

Extensive evaluations on three benchmarks well demonstrate the effectiveness of our SKT for extensive crowd counting models. In this project, the well-trained teacher networks and the distilled student networks have been released at GoogleDrive and BaiduYun (extract code: srpl). If you use this code and the released models for your research, please cite our paper:

@inproceedings{liu2020efficient,
  title={Efficient Crowd Counting via Structured Knowledge Transfer},
  author={Liu, Lingbo and Chen, Jiaqi and Wu, Hefeng and Chen, Tianshui and Li, Guanbin and Lin, Liang},
  booktitle={ACM International Conference on Multimedia},
  year={2020}
}

Datasets

ShanghaiTech: Google Drive

UCF-QNRF: Link

Prerequisites

We strongly recommend Anaconda as the environment.

Python: 2.7

PyTorch: 0.4.0

Preprocessing

1. Generation the ground-truth density maps for training

# ShanghaiTech
python preprocess/ShanghaiTech_GT_generation.py

# UCF-QNRF
python preprocess/UCF_GT_generation.py --mode train
python preprocess/UCF_GT_generation.py --mode test

2. Make data path files and edit this file to change the path to your original datasets.

python preprocess/make_json.py

Training

Edit this file for distillation training

bash SKT_distill.sh

Testing

Edit this file for testing models

bash test.sh

Models

The well-trained teacher networks and the distilled student networks are released at have been released at GoogleDrive and BaiduYun (extract code: srpl ). In particular, only using around 6% of the parameters and computation cost of original models, our distilled VGG-based models obtain at least 6.5× speed-up on an Nvidia 1080 GPU and even achieve state-of-the-art performance.

Shanghaitech A (576×864)

Method	MAE	RMSE	#Param (M)	FLOPs (G)	GPU (ms)	CPU (s)	Comment
CSRNet	68.43	105.99	16.26	205.88	66.58	7.85	teacher model, trained with CSRNet-pytorch
1/4-CSRNet + SKT	71.55	114.40	1.02	13.09	8.88	0.87	--
BL	61.46	103.17	21.50	205.32	47.89	8.84	teacher model
1/4-BL + SKT	62.73	102.33	1.35	13.06	7.40	0.88	--

UCF-QNRF (2032×2912)

Method	MAE	RMSE	#Param (M)	FLOPs (G)	GPU (ms)	CPU (s)	Comment
CSRNet	145.54	233.32	16.26	2447.91	823.84	119.67	teacher model, trained with CSRNet-pytorch
1/4-CSRNet + SKT	144.36	234.64	1.02	155.69	106.08	9.71	--
BL	87.70	158.09	21.50	2441.23	595.72	130.76	teacher model
1/4-BL + SKT	96.24	156.82	1.35	155.30	90.96	9.78	The released model is much better.