Choose a route for your scene根据场景选择入口
Region supplied: estimate pose from each instance's mesh and mask or box. Region to find: detect regions first, then estimate pose; supplied categories and counts define localization, while unknown counts use detection. For video, initialize by either route, then update the known tracks together. Category means a CAD identity; instance means a particular object in the scene.
已有物体区域：根据每个实例的网格与 mask 或包围盒估计位姿。需要寻找物体区域：先检测区域，再估计位姿；指定类别与数量时按定位任务使用，数量未知时按检测任务使用。视频先通过其中一条路径初始化，再批量更新已建立的轨迹。类别是 CAD 型号，实例是场景中的具体物体。
| Categories and instances类别与实例 | Region supplied已有区域：估计 | Region to find寻找区域：定位／检测 | Video update视频：跟踪更新 |
|---|---|---|---|
| One category, one instance单类、单实例 | Estimate one supplied region估计一个指定区域 | Search one category; retain at most one pose搜索一类，最多保留一个位姿 | Update one track更新一条轨迹 |
| One category, several instances单类、多实例 | Batch regions sharing a mesh同一网格、多个区域批量计算 | Search one category; supply a count or leave it uncapped搜索一类，指定数量或不限制数量 | Batch distinct track IDs sharing a mesh同类不同轨迹 ID 批量计算 |
| Several categories, one instance each多类、每类单实例 | Batch regions with their respective meshes各区域携带对应网格批量计算 | Search several categories; retain at most one pose per category搜索多类，每类最多保留一个位姿 | Batch one track per category每类一条轨迹批量计算 |
| Several categories, several instances each多类、每类多实例 | Batch all supplied instances全部指定实例批量计算 | Supply per-category counts or leave them uncapped按类别指定数量或不限制数量 | Batch all tracks, retaining their IDs全部轨迹批量计算，保留各自 ID |
Use estimate_one_category_one_instance for one supplied region, estimate_many_categories_many_instances for a region batch, estimate_frame_many_categories_many_instances when detection is needed, and track_many_categories_many_instances for known track updates. The “many” entries accept a single category or instance too. Tracking batches WAPR/SAPR and WBPS across instances, with independent score groups; engine limits split a large list into chunks. Track association and lost-object recovery are separate caller responsibilities.
一个指定区域使用 estimate_one_category_one_instance；区域批次使用 estimate_many_categories_many_instances；需要检测时使用 estimate_frame_many_categories_many_instances；已有轨迹更新使用 track_many_categories_many_instances。名字含“many”的入口也接受单类或单实例。跟踪的 WAPR/SAPR 与 WBPS 跨实例批量计算，物体评分组相互独立；大批次按引擎容量分块。轨迹关联与丢失恢复由调用方另外处理。
A one-ID obj_ids list restricts detection to that category but can still return several instances. In BOP localization, the present identities and counts are supplied; in detection, they are not. Here “unseen” means absent from pose-model training, with a mesh supplied at inference time; these application examples have a different object-preparation procedure from BOP's official model-free evaluation. Neither a 2D score nor a WBPS score is a calibrated success probability.
obj_ids 只给一个 ID，会将检测限定在该类别，但仍可能返回多个实例。BOP 的定位任务提供当前图像中的物体类别与数量，检测任务则不提供。这里的“未见”是指位姿模型训练时没有该物体、推理时提供网格；这些应用示例的物体准备方式与 BOP 官方无模型评测不同。2D 分数和 WBPS 分数都不是校准过的成功概率。