SCAR: Spatial-/Channel-wise Attention Regression Networks for Crowd Counting (1908.03716v1)

Abstract: Recently, crowd counting is a hot topic in crowd analysis. Many CNN-based counting algorithms attain good performance. However, these methods only focus on the local appearance features of crowd scenes but ignore the large-range pixel-wise contextual and crowd attention information. To remedy the above problems, in this paper, we introduce the Spatial-/Channel-wise Attention Models into the traditional Regression CNN to estimate the density map, which is named as "SCAR". It consists of two modules, namely Spatial-wise Attention Model (SAM) and Channel-wise Attention Model (CAM). The former can encode the pixel-wise context of the entire image to more accurately predict density maps at the pixel level. The latter attempts to extract more discriminative features among different channels, which aids model to pay attention to the head region, the core of crowd scenes. Intuitively, CAM alleviates the mistaken estimation for background regions. Finally, two types of attention information and traditional CNN's feature maps are integrated by a concatenation operation. Furthermore, the extensive experiments are conducted on four popular datasets, Shanghai Tech Part A/B, GCC, and UCF_CC_50 Dataset. The results show that the proposed method achieves state-of-the-art results.

• The paper introduces a novel dual attention mechanism integrating spatial and channel modules to enhance feature extraction in crowd counting.
• The paper demonstrates state-of-the-art performance on four major datasets by significantly reducing MAE and MSE compared to existing methods.
• The paper suggests that its flexible attention modules can be adapted for other pixel-wise prediction tasks, broadening the potential impact on computer vision.

SCAR: Spatial-/Channel-wise Attention Regression Networks for Crowd Counting

The paper presents the development of SCAR (Spatial-/Channel-wise Attention Regression Networks) tailored for the task of crowd counting in computer vision. Crowd counting is a crucial task for urban planning, public safety, and resource management, where accurate predictions of people densities in crowd scenes are vital. Traditional CNN-based methods, while effective, often focus only on local features and neglect broader contextual and channel-specific information, thereby limiting their predictive accuracy.

SCAR innovatively incorporates attention mechanisms into the traditional regression CNN framework, specifically leveraging Spatial-wise and Channel-wise Attention Models (SAM and CAM). These modules are designed to enhance feature extraction in two complementary ways. SAM encodes pixel-wise spatial contexts to ensure accurate positioning within density maps, directly addressing the limitations in capturing large-range spatial contextual information. CAM, on the other hand, addresses challenges related to channel dependencies by enhancing feature discrimination across channels, reducing erroneous crowd region predictions in background-rich scenes.

The paper provides an extensive validation of SCAR's performance across four prominent datasets: Shanghai Tech Part A/B, GCC, and UCF_CC_50. Results indicate that SCAR achieves state-of-the-art accuracy in density map prediction, outperforming existing approaches such as MCNN and CSRNet in terms of MAE and MSE metrics. The integration of SAM and CAM not only improves counting accuracy but also enhances the quality of density maps as evaluated by PSNR and SSIM metrics.

The theoretical contribution of this work lies in its dual attention strategy that combines spatial and channel information, providing a more holistic understanding of crowded scenes. This approach effectively bridges the gap between local feature extraction and the need for global contextual awareness in dense environments. Practically, SCAR offers significant potential for real-time applications in surveillance systems where accurate crowd estimation is paramount.

For future research, the paper suggests that the attention mechanisms proposed in SCAR could be adapted for other pixel-wise prediction tasks such as image segmentation and saliency detection. The flexibility of SCAR's attention modules means that its principles can easily be transferred to other domains within the field of computer vision, offering new avenues for enhancing context-aware feature extraction in a variety of visual analysis tasks.