HOME CHALLENGES DATASETS TASKS LEADERBOARDS SUBMIT RESULTS Sign in

Submission: SAM-FP/XYZ-IBD/CyclicFP

Download submission
Submission name CyclicFP
Submission time (UTC) Sept. 11, 2026, 1:35 a.m.
User yizhouzhao
Task Model-based 6D detection of unseen objects
Dataset XYZ-IBD
Description Per image: depth is taken from the xyz sensor's own map (no depth prediction). For each of the 15 classes, SAM3 produces mask proposals from one short text prompt, and FoundationPose estimates a pose for every (class, proposal) pair using the provided mesh. Proposals are ranked by the FoundationPose score network alone; the pipeline's other rerank terms (SAM score, rendered mask IoU, rendered box IoU) are weighted zero.
Evaluation scores
AP:0.765
AP_25:0.776
AP_25_mm:0.645
AP_MSPD:0.753
AP_MSSD:0.776
AP_MSSD_mm:0.645
average_time_per_image:387.284

Method: SAM-FP

User yizhouzhao
Publication
Implementation Python pipeline built on PyTorch (SAM3) and TensorRT (FoundationPose), run on Ubuntu 24.04, CUDA 13.2
Training image modalities RGB-D
Test image modalities RGB-D
Description

Method: SAM3 + FoundationPose (single-view RGB-D).

A training-free pipeline combining two published, off-the-shelf models. No network was trained or fine-tuned; all weights are used as released, and the same models handle all objects without per-object training.

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

Pipeline (per BOP24 target image):

  1. Input — a single target RGB-D view. The provided sensor depth is used directly; no additional views or depth reconstruction are used.
  2. Detection and segmentation (SAM3) — text-prompted, open-vocabulary detection on the target RGB image, with a confidence threshold of 0.10.
  3. 6D pose estimation (FoundationPose) — model-based registration using the target RGB image, sensor depth, and provided CAD models. FoundationPose uses 128 pose hypotheses and five refinement iterations. All pose proposals are submitted without a score cutoff; the score column contains FoundationPose’s reranking score.

Differences with respect to the linked publications:

  • SAM3 and FoundationPose are composed into an integration pipeline.
  • The individual networks are otherwise unmodified from their released versions.

Training images and rendering: None. No training or offline rendering was performed. During inference, FoundationPose renders CAD-model hypotheses online for pose refinement and scoring. The provided models_eval CAD meshes are used directly.

Networks: One set of networks is used for all objects—a single SAM3 model and a single FoundationPose model. No per-object models are used.
Computer specifications L40s server