Publication 04 · Crowd counting
Attention Mixture Network via Binarization Transfer.
Two kinds of attention answer two different questions — where people are, and how dense it is. AMNet refuses to pick one.
Binary attention maps tell a crowd counter where the crowd is; non-binary maps capture the density gradient inside it. Prior work used one or the other. AMNet feeds shared features to both an attention-map generator and a density-map generator, and couples them through a Binarization Transfer Module (BTM) that jointly optimizes the binarization during training — filtering cluttered backgrounds while keeping density structure.
The question
Crowd counting lives or dies on attention. The binary kind separates crowd regions from intricate background; the non-binary kind characterizes congestion level inside those regions. Methods historically commit to one — and pay for it either in background noise or in lost density structure.
The question: can both maps be exploited at once, with the binarization itself learned rather than hand-thresholded?
The approach
AMNet shares a feature extractor between an attention map generator (binary, candidate crowd regions) and a density map generator (non-binary, congestion gradients). The Binarization Transfer Module turns binarization into a differentiable, jointly-optimized step — transferring the clean region selection from the binary map into the density estimation while preserving multi-scale detail.

What it shows
Across ShanghaiTech PartA and PartB, UCF_CC_50, WorldExpo'10, and UCF-QNRF, AMNet achieves significant improvements in counting accuracy and outperforms state-of-the-art methods — with the largest gains exactly where binary attention alone fails: dense, cluttered scenes.
Qualitative density maps show the mechanism working as designed: backgrounds suppressed by the binarized attention while density gradients inside the crowd remain intact.

Claim boundary
✓ Established
- Simultaneous use of binary and non-binary attention maps
- A differentiable Binarization Transfer Module for joint training
- Improvements across four standard crowd-counting benchmarks
- Qualitative evidence of background suppression with preserved density
✗ Explicitly not claimed
- Benchmark scope: the five standard counting sets tested
- No deployment or real-time claim
- Privacy context of crowd imagery acknowledged, not studied
- No cross-domain (non-crowd) transfer claim
Citation
@inproceedings{xu2024amnet,
title = {Attention Mixture Network for Crowd Counting
via Binarization Transfer},
author = {Xu, Junjie and Zhang, Zihao and Li, Xin and Li, Weijie
and Yu, Kun},
booktitle = {Proceedings of the 2nd International Workshop on Multimedia
Content Generation and Evaluation (McGE)},
series = {ACM Multimedia 2024},
year = {2024},
doi = {10.1145/3688867.3690172}
}