2D detection and segmentation2D 检测与分割

WAPRDet2D.detect_many_categories_many_instances performs CAD-conditioned detection and visible-region segmentation from one RGB image. Preparation builds a reusable template bank by rendering CAD meshes and encoding the views with DINOv2^[2]. Per-frame inference then uses three model stages:

WAPRDet2D.detect_many_categories_many_instances 以单张 RGB 图像为输入，完成基于 CAD 的检测及可见区域分割。准备阶段渲染 CAD 网格，并用 DINOv2^[2] 编码视图，建立可复用的模板库。逐帧推理包含三个模型阶段：

1. GroundingDINO^[1] generates candidate bounding boxes.GroundingDINO^[1] 生成候选包围盒。
2. SAM segments candidate regions into visible masks.SAM 分割候选区域，生成可见区域掩码。
3. DINOv2 encodes the segmented regions and assigns CAD identities through template matching.DINOv2 编码分割区域，通过模板匹配确定 CAD 身份。

Each returned instance contains a detection box and a visible mask. The method does not use frame-level ground-truth object identities, instance counts, boxes, masks, or depth. Its score_2d ranks 2D detections and is distinct from the pose-level score_6d. The API calls its score gate confidence for compatibility; this fused matching score is not a calibrated probability.

每条返回结果均包含检测框和可见区域掩码。该方法不使用当前帧的真值物体类别、实例数量、包围盒、掩码或深度。返回的 score_2d 用于排序 2D 检测结果，与位姿层面的 score_6d 含义不同。接口沿用 confidence 作为分数阈值参数名；融合后的匹配分数并非校准概率。

• CAD templates and feature encodingCAD 模板与特征编码
• Detector usage and evaluation检测器使用与评估