About

Evaluates sampling-based Bayesian inference (different variants of MCMC, SMC) for the 6D pose estimation of objects using depth images and CAD models only. This code has been produced during while writing my Ph.D. (Dr.-Ing.) thesis at the Institut of Automatic Control, RWTH Aachen University. If you find it helpful for your research please cite this:

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.

I submitted my results of the best performing SMC sampler to the BOP benchmark with two different time budgets per pose estimate:

• 1 second: https://bop.felk.cvut.cz/method_info/458/
• 0.5 seconds: https://bop.felk.cvut.cz/method_info/457/

Setup

VSCode Devcontainer (recommended)

The easiest way to get started is using the VSCode Remote - Containers extension. This setup includes:

• startup.jl: automatically loads Julia packages ImageShow to visualize results in VSCode, Revise automatically recompiles code on changes, and OhMyREPL for an improved REPL experience.
• OpenGL: the docker-compose.yml includes all necessary mounts to run GUIs and OpenGL applications. You can use glxinfo -B to check which OpenGL driver is used.
• CUDA: the Dockerfile is based on the official CUDA image. Note: OpenGl-CUDA interop is not supported on WSL yet.

This will install all required dependencies and launch a Docker container which has access to the GPU (CUDA, DISPLAY, & OpenGL).

Manual Setup

Like any other Julia project, you can install the required packages using the Project.toml and Manifest.toml files via import Pkg; Pkg.activate("."); Pkg.resolve(); Pkg.instantiate(). I have tested the versions in the Manifest.toml file, but you can update them to the latest version since I have included github dependencies in the [sources] section of Project.toml:

• https://github.com/rwth-irt/BayesNet.jl - Type stable implementation of a Bayesian network.
• https://github.com/rwth-irt/KernelDistributions.jl - Subset of Distributions.jl which can be used in CUDA kernels.
• https://github.com/rwth-irt/PoseErrors.jl - 6D pose error metrics from BOP Challenge
• https://github.com/rwth-irt/SciGL.jl - Efficient rendering in OpenGL and CUDA interop for julia
• https://github.com/rwth-irt/BlenderProc.DissTimRedick - BlenderProc setup to generate the synthetic datasets from my dissertation.

Obtaining the Datasets

With the exception of the STERI dataset, all datasets can be obtained from the BOP benchmark. The STERI dataset contains proprietary CAD models, which I cannot make available.

My own data sets can be obtained from me [RWTH publications soon](TODO ongoing request).

• BOP Datasets includes the synthetically generated training datasets generated with BlenderProc. These can be used for parameter tuning.
• rosbags contains the rosbags to evaluate the particle filter. Ground truth data has been obtained using the forward kinematics of the robot holding the camera.

Extract the datasets into the data/bop directory, e.g. data/bop/itodd.

Notes on the Project Structure

I use DrWatson.jl to manage the experiments and follow their project setup conventions. For example, the data, scripts, and src folders are the result of this convention.

Note on recent CUDA.jl versions

v5.0.0 introduced some changes which negatively impact performance

• when assigning prior_o[mask_img] .= ... an "attempt to release free memory error occurs"
• benchmark simple_posterior vs. smooth_posterior after upgrading