CLAHR: Cascaded Label Assignment Head for High-Resolution Small Object Detection

• Published in: IEEE access Vol. 12; pp. 15447 - 15457
• Main Authors: Qingyong, Yang, Chenchen, Huang, Likun, Cao, Qi, Song, Xiyan, Jiang, Ximei, Liu, Chunmiao, Yuan
• Format: Journal Article
• Language: English
• Published: Piscataway The Institute of Electrical and Electronics Engineers, Inc. (IEEE) 2024; IEEE
• Subjects: Cascade structure; Detectors; label assignment; Labels; Normal distribution; Object recognition; Sampling; Sensors; small object detection
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3357984

Small object detection is one of the main obstacles hindering the development of object detection technology. In small object detection tasks, the performance of universal object detectors often drops sharply. We find through experiments on common models that for small objects, the box prior based on anchor detectors and the point prior without anchor detectors are suboptimal. The current anchor based or anchor-free label allocation methods will generate many false positive small objects samples, resulting in a decrease in detector attention to small objects. Therefore, we propose a new detector CLAHR for small object detection. In response to the issue of high sensitivity to small objects using Intersection over Union (IoU), where small deviations can lead to poor quality of assigned positive and negative samples, CLAHR utilizes the insensitivity of Gaussian distribution to small perturbations of small objects to transform the predicted box and manually annotated true box into Gaussian distribution. Then, CLAHR uses Wasserstein Distance (WD) to measure the distance between the predicted actual receptive field (ARF) and the manually labeled true receptive field (TRF) of the real box, rather than using IoU or central sampling strategies to allocate samples. Considering that IoU threshold based and central sampling strategies tend to favor large objects, we further design a cascaded label assignment (CLA) module to achieve balanced learning for small objects samples. Extensive experiments on the COCO, AI-TOD, and VisDrone datasets demonstrate the effective-eness of the proposed approach.