Unverified60% confidenceFactExact time
Two-stage检测器（如R-CNN、Faster R-CNN）先通过区域提议网络（RPN）生成候选框再逐一分类，典型推理速度仅在5-10 FPS左右
2
Sources
60%
Confidence
Long-term
Relevance
7/8/2026
First Seen
Sources
YOLO入门学习路径：版本演进理解到源码精读实战指南
bilibili迪哥带你学AI7/7/2026
Related Claims
Unverified传统两阶段方法如Faster R-CNN依赖区域提议网络（RPN）先生成候选框，推理延迟高达100ms以上83% similarUnverifiedTwo-stage检测器先通过RPN生成候选框再逐一分类，精度高但速度慢；One-stage检测器将候选框生成与分类合并为一步，牺牲少量精度换取数十倍速度提升81% similarUnverifiedRLHF微调阶段相比预训练在数据量级上通常小2-3个数量级，使安全行为编码为低秩覆写而非底层重塑66% similarUnverified增加重排序步骤即可将检索准确率（NDCG@10）提升 10-20 个百分点64% similarUnverified为每个传感器独立建模测量协方差矩阵R是多传感器融合区别于单传感器扩展的本质，R设置过小会导致过度信任该传感器测量值64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/276032API
curl https://kongchang.com/api/v1/knowledge/claims/276032MCP
get_claim(id=276032)