TWO-STAGE CLASSIFICATION ARCHITECTURES IN MODERN OBJECT DETECTION: FROM REGION PROPOSALS TO FINE-GRAINED RECOGNITION
Keywords:
CLASSIFICATION, ARCHITECTURES, MODERN, single, application, regionAbstract
Object detection architectures are broadly divided into single-stage and two-stage families. Single-stage detectors, such as the YOLO and SSD lineages, predict bounding boxes and class labels in a single dense pass over the image, prioritizing speed. Two-stage detectors, beginning with R-CNN and refined through Fast R-CNN, Faster R-CNN, and their many descendants, separate the task into a region proposal stage followed by a classification and refinement stage, generally trading some inference speed for improved localization and classification accuracy.This article examines the architectural principles behind two-stage and cascaded classification pipelines, independent of any single application domain, to clarify why splitting detection into sequential stages has proven so durable as a design pattern. The discussion moves from the basic rationale for splitting localization and classification, through the internal anatomy of the second-stage classifier, to the specific architectural refinements — cascading, head decoupling, and attention — that have been proposed to strengthen it.
References
1.Girshick, R. (2015). Fast R-CNN.
2.Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks.
3.He, K., Gkioxari, G., Dollár, P., & Girshick, R. (2017). Mask R-CNN (introduced RoI Align).
4.Cai, Z., & Vasconcelos, N. (2018). Cascade R-CNN: Delving into High Quality Object Detection.
5.Wu, Y., Chen, Y., Yuan, L., Liu, Z., Wang, L., Li, H., & Fu, Y. (2020). Rethinking Classification and Localization for Object Detection (Double-Head R-CNN).
6.Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., & Zagoruyko, S. (2020). End-to-End Object Detection with Transformers (DETR).