Reproduce “Encoder-Decoder Based Convolutional Neural Networks with Multi-Scale-Aware Modules for Crowd Counting” on OpenResearch

About this paper

Pongpisit Thanasutives, Ken-ichi Fukui, Masayuki Numao, Boonserm Kijsirikul

This research introduces M-SFANet and M-SegNet, enhanced encoder-decoder convolutional neural networks that effectively tackle scale variation and background noise in crowd counting. M-SFANet achieves state-of-the-art accuracy across multiple datasets by integrating adaptive multi-scale modules, while M-SegNet offers a computationally efficient alternative with competitive performance through max unpooling.

Key insights

• Combining adaptive (CAN) and static (ASPP) multi-scale modules within an encoder-decoder framework proves effective for robust scale variation handling across diverse crowd densities.
• Max unpooling, a technique from semantic segmentation, can be successfully adapted for crowd counting, offering a computationally efficient upsampling method that maintains competitive accuracy.
• The dual-path decoder, producing both density and attention maps, is crucial for effectively suppressing background noise and improving the reliability of crowd density estimation.

Summary by alphaXiv.