WAPR.v2 adopts the same zero-shot 2D detector setting as FRTPose-WAPR.v2. Unlike FRTPose-WAPR.v2, WAPR.v2 removes the FRTPose component, making it a fully zero-shot method. This modification increases inference speed but results in a slight reduction in accuracy. Similar to FRTPose-WAPR.v2, WAPR.v2 initializes 24 uniformly sampled candidate poses for each detected object and applies the WAPR module to refine each pose five times. The WAPR module supports wide-angle pose refinement, allowing initialization errors of up to ±90°. Finally, the refined poses are scored by the FoundationPose pose scoring network, and the pose with the highest score is selected as the final result.

For a task involving 30 different objects and an image containing 100 detected 2D bounding boxes, the total computational workload can be expressed as 30 (objects) × 100 (detections) × 24 (candidate poses) × (5 (refinements) + 1 (scoring)) = 432,000. These 432,000 parallel operations represent the overall computation scale rather than sequential iterations. Under this configuration, the average computation time per image corresponds to the time required for handling this parallel workload. If the 2D detection confidence scores and classifications are considered reliable, a coarse filtering of low-confidence bounding boxes can be applied. This significantly reduces the number of candidate detections, thereby reducing the average computation time per image to less than one second.

In addition, to further reduce computation time, non-maximum suppression (NMS) is applied to remove redundant 2D bounding boxes. This reduces the overall runtime from approximately 120 s to approximately 50 s.