AlignDet is an innovative pre-training framework developed to address discrepancies in data, model, and task between pre-training and fine-tuning stages in object detection algorithms. These disparities potentially limit a detector’s performance, generalization ability, and convergence speed. AlignDet alleviates these issues by decoupling the pre-training process into image-domain and box-domain stages, optimizing the detection backbone and learning instance-level semantics, respectively. As a result, AlignDet yields significant improvements in performance across diverse protocols, such as detection algorithms, model backbones, data settings, and training schedules.
Publication date: July 20, 2023
Project Page: https://liming-ai.github.io/AlignDet
Paper: https://arxiv.org/pdf/2307.11077.pdf
