Efficient and Accurate Crowd Counting with Lightweight Models

In this article, researchers propose a new approach to crowd counting that takes into account the scale of the scene. Traditional methods rely on pixel-wise measurements, which can lead to inaccurate results when dealing with crowds at different distances from the camera. The proposed method, called Scale-Aware Crowd Counting Network (SACC-Net), uses multifaceted attention to focus on different scales of the scene and estimate the crowd density accordingly.
The SACC-Net architecture consists of several parts: a dense attention network that generates various attention masks based on the scale of the scene; a scale-aware loss function that takes into account the size of each object in the scene; and an intra-block fusion module to allow all feature layers within the same convolution block to be fused, allowing for more fine-grained information to be sent to the decoder.
The proposed method outperforms traditional methods on four popular crowd counting datasets, demonstrating its effectiveness in handling crowds at different distances from the camera. The researchers also propose a novel synthetic fusion module (SFM) to scale the feature maps to different scales, and an intra-block fusion module (IFM) to allow all feature layers within the same convolution block to be fused.
The key innovation of SACC-Net is its ability to handle crowds at different distances from the camera by using multifaceted attention to focus on different scales of the scene. This allows the method to accurately estimate crowd density in a wide range of scenarios, including scenes with varying object sizes and distances from the camera.
In summary, SACC-Net is a scale-aware crowd counting network that uses multifaceted attention to handle crowds at different distances from the camera. By using a novel fusion module to scale the feature maps and an intra-block fusion module to allow all feature layers within the same convolution block to be fused, SACC-Net outperforms traditional methods on four popular crowd counting datasets.