Submission: SMC-1.0s-CNOS/HB

Download submission
Submission name
Submission time (UTC) Nov. 12, 2023, 10:50 a.m.
User Tuebel
Task Model-based 6D localization of unseen objects
Dataset HB
Training model type Default
Training image type None
Description
Evaluation scores
AR:0.585
AR_MSPD:0.587
AR_MSSD:0.583
AR_VSD:0.585
average_time_per_image:5.614

Method: SMC-1.0s-CNOS

User Tuebel
Publication
Implementation
Training image modalities None
Test image modalities D
Description

Submitted to: BOP Challenge 2023

Training data: None

Onboarding data: None

Used 3D models: CAD or default

Notes:

The localization-only method requires a separate detector and is only tested on Task 4 with default segmentation masks from CNOS. This version runs the pose inference at most for inst_count instances from test_targets_bop19.json.

It uses a sequential Monte Carlo (SMC) sampler and the CAD model to estimate the pose using only the segmentation mask and the depth image. The GPU is utilized to render depth images for pose hypotheses and evaluate their likelihood in parallel.

The sampler had a time budget of 1.0s per object pose inference.

Part of my dissertation: T. Redick, „Bayesian inference for CAD-based pose estimation on depth images for robotic manipulation“, RWTH Aachen University, 2024. doi: 10.18154/RWTH-2024-04533. Code: https://github.com/rwth-irt/BayesianPoseEstimation.jl

Contact: tim.redick@rwth-aachen.de

Computer specifications AMD Ryzen Threadripper PRO 5975WX, NVIDIA RTX 4090