Method: SAM3 + FoundationStereo + FoundationPose (multi-view RGB, reconstructed depth).

A training-free pipeline combining three published, off-the-shelf models — no network was trained or fine-tuned; all weights are used as released, and one model handles all objects (no per-object training).

Codebase: https://github.com/nvidia-isaac/foundationpose_perception_pipeline

Pipeline (per BOP19 target image): 1. Depth (FoundationStereo, research checkpoint 23-51-11) — sensor depth is not used. Depth is reconstructed by treating the target frame plus 1–6 additional test frames of the same static scene (camera baseline 0.10–0.24 m) as a stereo pair. This makes the method multi-view RGB.

1. Detection/segmentation (SAM3) — text-prompted open-vocabulary detection on the single target RGB view (confidence threshold 0.10).

2. 6D pose (FoundationPose) — model-based registration on the target RGB + reconstructed depth using the provided CAD models (128 pose hypotheses, 5 refinement iterations). All posed proposals are submitted (no score cutoff); the score column is FoundationPose's rerank score.

Differences w.r.t. the linked publications: - Depth is not from the RGB-D sensor but reconstructed via multi-view stereo over the static test scene, so the method is multi-view RGB rather than RGB-D. - The FoundationStereo research checkpoint runs on Blackwell (sm_120) with torch 2.10/cu128 instead of the paper's torch 2.4.1/cu121 build. - Models are composed as an integration pipeline; the individual networks are unmodified from their releases.

Training images / rendering: none — no training or rendering was performed. FoundationPose renders CAD model hypotheses online at inference time for pose refinement/scoring; the provided models_eval CAD meshes are used directly.

Networks: one set of networks for all objects (a single SAM3, one FoundationStereo, one FoundationPose); no per-object models.