检测技术¶
通用目标检测
- RefineDet
- "Single-Shot Refinement Neural Network for Object Detection"
- 推荐,一个二次回归的单段检测器,相对于 SSD,速度下降不大,但准确率显著提升。
- Detect at 200 FPS
- "Object detection at 200 Frames Per Second"
- 推荐,一个二次回归的单段检测器,相对于
人脸检测
- SRN: Seletive Refinement Network
- "Selective Refinement Network for High Performance Face Detection"
- 推荐,根据 RefineDet 思路发展而来的人脸检测器,效果很好。
- DSFD: Dual Shot Face Detector
- "DSFD: Dual Shot Face Detector"
- 不推荐,虽然目前性能最好,但提升主要来自于使用复杂网络。
- FaceBoxes: SSD-style 的快速检测器
- "FaceBoxes: A CPU Real-time Face Detector with High Accuracy"
- 最小脸 20px,20FPS@CPU,125FPS@GPU。
- DCFPN:Single Head 快速检测器
- "Detecting Face with Densely Connected Face Proposal Network"
- 最小脸 40px,30FPS@CPU, 250FPS@GPU。
目录
Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference