Context Enhancement Module for Object Detection
In object detection, the Context Enhancement Module (CEM) is a feature extraction module used specifically in ThunderNet which enlarges the receptive field. The aim of the CEM is to aggregate multi-scale local context information and global context information to generate more discriminative features.
The Key Concepts of CEM
CEM is designed to merge feature maps from three scales - C4, C5, and Cglb. Cglb is the global context feature vector obtained by using global average pooling on C5. One convolution of 1x1 is applied on each feature map to squeeze the number of channels to α x p x p=245.
After this step, C5 is upsampled by 2x, and Cglb is broadcast so that the spatial dimensions of the three feature maps are equal. Finally, the three generated feature maps are aggregated. By leveraging both local and global context, CEM effectively enlarges the receptive field and refines the representation ability of the thin feature map.
Advantages of CEM over Prior FPN Structures
Compared with prior FPN structures, CEM involves only two 1x1 convolutions and a fc layer. This makes it computationally efficient and more accurate. Additionally, CEM can help to improve detection accuracy in a variety of different scenarios, such as remote sensing imagery, medical brain scans, and more. Overall, CEM provides an effective way to improve the representation ability of feature maps, and hence improve the overall performance of object detectors.
By merging multi-scale local context information and global context information, Context Enhancement Module (CEM) generates more discerning features in object detection. CEM not only enlarges the receptive field but it also provides an effective way to refine the representation ability of feature maps. Compared to prior FPN structures, CEM is computationally efficient and can improve detection accuracy in numerous fields.
