ObjectPoseEstimation

Methods for zero-shot 6D object pose estimation from RGB(-D) images:

• ZS6D
• SAM-6D
• FoundPose

Evaluation on the BOP datasets.

Requirements

2025/05/20 - Tested with:

• 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz - 1 socket, 8 cores per socket, 2 threads per core
• 32GiB RAM - 2 x 16GiB SODIMM DDR4 Synchronous 3200 MHz
• NVIDIA GeForce RTX 3080 Mobile 16GB
• Ubuntu 22.04.5
• NVIDIA Driver Version: 535.247.01
• Docker version 28.1.1, build 4eba377
• NVIDIA Container Toolkit 1.17.6
• OSRF/rocker 0.2.19

Installation

Clone the repository:

git clone https://github.com/RoboticRepositories/ObjectPoseEstimation.git
cd ObjectPoseEstimation && git submodule update --init --recursive

Note: some of the submodules use SSH URLs. For cloning them properly you must generate an SSH keypair on your computer and add the public key to your account on GitHub. For more information, see Connecting to GitHub with SSH.

Datasets

./Datasets/bop/detections/download.sh

./Datasets/bop/lmo/download.sh

./Datasets/bop/ycbv/download.sh

./Datasets/bop/tless/download.sh

ZS6D

Build the Docker image:

./Docker/ZS6D/build.sh

Or pull it from Docker hub:

./Docker/ZS6D/pull.sh

Run a Docker container:

./Docker/ZS6D/run.sh 

Jupyter is running in the Docker container:

1. Download templates for YCBV
2. Prepare the templates
3. Run an inference test

SAM-6D

Build the Docker image:

./Docker/SAM-6D/build.sh

Or pull it from Docker hub:

./Docker/SAM-6D/pull.sh

Run a Docker container:

./Docker/SAM-6D/run.sh 

Jupyter is running in the Docker container:

1. Download models
2. LM-O test
   1. Render templates
   2. Run an inference test

3. YCBV test
   1. Render templates
   2. Run an inference test

4. TLESS test
   1. Render templates
   2. Run instance segmentation model
   3. Run pose estimation model

FoundPose

Build the Docker image:

./Docker/FoundPose/build.sh

Or pull it from Docker hub:

./Docker/FoundPose/pull.sh

Run a Docker container:

./Docker/FoundPose/run.sh 

Jupyter is running in the Docker container:

1. Download templates for LM-O
2. Generate object representations for LM-O
3. Run an inference test
   1. Script for all the images and objects
   2. Notebook for a single object in an image