Matrix Nets: A New Deep Architecture for Object Detection
-
AP
在MS COCO上实现了47.8的mAP -
AP比较

-
Matrix Networks

(a) Shows the original FPN architecture
(b) Shows the MatrixNet architecture, where the 5 FPN layers are viewed as the diagonal layers in the matrix. We fill in the rest of the matrix by downsampling these layers. -
KP-xNet architecture

-
实验结果
Empirical results on MSCOCO for our method as compared to best results reported in other works.
-
其他详细细节可以参考原论文
-
论文链接:Matrix Nets: A New Deep Architecture for Object Detection
-
开源代码
pytorch版 MatrixNets 实现: https://github.com/arashwan/matrixnet (包含了 train、 test 代码)

MatrixNets是一种新的深度学习架构,专为物体检测设计,在MSCOCO数据集上实现了47.8的mAP,超越了传统的FPN架构。通过在矩阵中填充下采样层,形成了一种创新的网络结构。
2844

被折叠的 条评论
为什么被折叠?



