CenterNet-Elite: A Small Object Detection Model for Driving Scenario

With the rapid development of deep learning networks, the accuracy of generic object detection has consistently improved. Nonetheless, small object detection tasks still face a range of challenges. On one hand, the limited pixel size of small objects severely constrains their visual features in images, making them susceptible to distortion and difficult to distinguish from background noise. On the other hand, small objects often appear in complex scenarios with severe occlusion and dense arrangement, further increasing the complexity and difficulty of small object detection tasks. In this context, this study proposes a new model, CenterNet-Elite, to overcome these challenges. To address the issue of feature information loss resulting from multiple downsamplings during feature extraction for small objects, we introduce the spatial and channel reconstruction convolution (SCConv) into the bottleneck to reduce spatial and channel redundancy and enhance feature representation. In the meantime, we construct multiple short connections to integrate feature maps of the same scale during the downsampling and upsampling stages, thereby retaining critical shallow spatial information. We introduce a multi-scale pooling module, SPPCSPC, to address the challenge of significant variations in object scale. This module obtains receptive fields of different sizes through max-pooling layers of diverse sizes, adapting to changes in object scale on the feature maps. Furthermore, we introduce the content-aware reassembly of features (CARAFE) to replace deconvolution, refining the upsampling process to enhance the quality of feature maps. A series of comparative experiments and ablation studies demonstrate the effectiveness of our method in small object detection. The CenterNet-Elite achieves a 2.3% increase in the average precision on the large-scale small object detection dataset (SODA-D).