STATION 09 · INSTANCE SEGMENTATION
实例分割：Instance Mask
展示 bbox、class、score、mask、instance id 的联合输出，解释 Mask R-CNN、ROI Align、Mask Head 与 YOLO-seg 推理。
INSTANCE MASK · YOLO-SEG
INSTANCE SEGMENTATION VISUALIZATION
Mask AP --
真实图像实例分割可视化
Source: Auto Model / Fallback
Model: --
Backend: --
Inference: --
Postprocess: --
Instances: --
ALGORITHM FLOW
逐步运算流程
图片未找到
ImageH×W×3
→
PreprocessLetterbox 640
→
InferenceYOLO11n-seg
→
Decodeboxes + coeffs
→
NMSfilter overlaps
→
Prototype32 masks blend
→
Instancescrop + stats
Model Feature Internals
模拟数据
YOLO11n-seg 原型特征融合演示仪 (Prototype Blender)
$$\text{Mask}(x,y)=\sigma\left(\sum_{i=1}^{32} C_i\times P_i(x,y)\right)$$
用神经网络提取的32个共享原型与各自的掩膜系数进行线性组合，求和后对其经过激活函数和BBox掩盖最终截断输出
活跃的原型通道 (Active Prototypes $P_{1..8}$)
正在等待提取通道特征...
- 1ImageCurrent / Uploaded
- 2PreprocessLetterbox + CHW
- 3Model InferenceONNX Runtime Web
- 4Decode Boxesscores + coeffs
- 5NMSIoU suppression
- 6Mask Prototypecoeff × proto
- 7Instance Maskscrop + statistics