Although large-scale pretraining followed by downstream fine-tuning is a prevalent approach in object detection, it often underperforms on datasets with significant long-tailed distributions. Our investigation identifies biases originating not only from extreme imbalances in classifier weight norms but also from simplicity biases at the feature representation level. To address these challenges, we introduce a novel pretraining methodology, Dynamic Rebalancing Contrastive Learning with Dual Reconstruction (DRCL). This method seamlessly integrates holistic and object-level contrasts within a contrastive learning framework, utilizes a dynamic rebalancing technique that transitions from image-level to instance-level resampling, and implements a dual reconstruction strategy to preserve both natural appearance and internal semantic consistency. By synergistically combining self-supervised and supervised learning modalities, our approach substantially reduces pretraining time and resource demands. Demonstrating significant enhancements over traditional long-tailed detection methods, particularly for rare classes, our methodology achieves State-of-the-Art performance on the extensive LVIS dataset across multiple detection frameworks and backbone networks.