bop-benchmark/datasets

BOP: Benchmark for 6D Object Pose Estimation

The goal of BOP is to capture the state of the art in estimating the 6D pose, i.e. 3D translation and 3D rotation, of rigid objects from RGB/RGB-D images. An accurate, fast, robust, scalable and easy-to-train method that solves this task will have a big impact in application fields such as robotics or augmented reality.

Homepage: https://bop.felk.cvut.cz/home/

Toolkit: https://github.com/thodan/bop_toolkit

Downloading datasets

Option 1: Using huggingface_hub:

Click to expand

a. Install the library:

pip install --upgrade huggingface_hub

b. Download the dataset:

from huggingface_hub import snapshot_download

dataset_name = "hope"
local_dir = "./datasets"

snapshot_download(repo_id="bop-benchmark/datasets", 
                  allow_patterns=f"{dataset_name}/*zip",
                  repo_type="dataset", 
                  local_dir=local_dir)

If you want to download the entire BOP datasets (~3TB), please remove the allow_patterns argument. More options are available in the official documentation.

Option 2: Using huggingface_hub[cli]:

Click to expand

a. Install the library:

pip install -U "huggingface_hub[cli]"

b. Download the dataset:

export LOCAL_DIR=./datasets
export DATASET_NAME=hope

huggingface-cli download bop-benchmark/datasets --include "$DATASET_NAME/*.zip" --local-dir $LOCAL_DIR --repo-type=dataset  

Please remove this argument --include "$DATASET_NAME/*.zip" to download entire BOP datasets (~3TB). More options are available in the official documentation.

Option 3: Using wget:

Click to expand

Similar wget command as in BOP website can be used to download the dataset from huggingface hub:

export SRC=https://huggingface.co/datasets/bop-benchmark/datasets/resolve/main

wget $SRC/lm/lm_base.zip         # Base archive 
wget $SRC/lm/lm_models.zip       # 3D object models
wget $SRC/lm/lm_test_all.zip     # All test images ("_bop19" for a subset)
wget $SRC/lm/lm_train_pbr.zip    # PBR training images 

Datasets are stored in .zip format. You can extract them using the following command:

bash scripts/extract_bop.sh

If you are running on a machine with high bandwidth, you can increase your download speed by adding the following environment variable:

pip install huggingface_hub[hf_transfer]
export HF_HUB_ENABLE_HF_TRANSFER=1

Uploading datasets

You create a new dataset and want to share it with BOP community. Here is a step-by-step guide to upload the dataset and create a pull request to our huggingface hub. Feel free to reach out to vanngn.nguyen@gmail.com if you have any questions.

Similar to the download process, you can upload the dataset using the huggingface_hub library or huggingface_hub[cli]. We recommend using huggingface_hub[cli] for its simplicity.

Option 1: Using huggingface_hub[cli]:

Click to expand

a. Install the library:

pip install -U "huggingface_hub[cli]"

b. Log-in and create a token

huggingface-cli login

Then go to this link and generate a token. IMPORTANT: the token should have write access as shown below:

Make sure you are in the bop-benchmark group by running:

huggingface-cli whoami

c. Upload dataset:

The command is applied for both folders and specific files:

# Usage:  huggingface-cli upload bop-benchmark/datasets [local_path] [path_in_repo] --repo-type=dataset --create-pr

For example, to upload hope dataset:

export LOCAL_FOLDER=./datasets/hope
export HF_FOLDER=/hope

huggingface-cli upload bop-benchmark/datasets $LOCAL_FOLDER $HF_FOLDER --repo-type=dataset --create-pr

Option 2: Using huggingface_hub:

Click to expand

a. Install the library:

pip install --upgrade huggingface_hub

b. Creating a pull-request:

We recommend organizing the dataset in a folder and then uploading it to the huggingface hub. For example, to upload lmo:

from huggingface_hub import HfApi, CommitOperationAdd

dataset_name = "lmo"
local_dir = "./datasets/lmo"

operations = []
for file in local_dir.glob("*"):
    add_commit = CommitOperationAdd(
        path_in_repo=f"/{dataset_name}",
        path_or_fileobj=local_dir,
    )
    operations.append(add_commit)


api = HfApi()
MY_TOKEN = # get from https://huggingface.co/settings/tokens
api.create_commit(repo_id="bop-benchmark/datasets", 
                  repo_type="dataset",
                  commit_message=f"adding {dataset_name} dataset", 
                  token=MY_TOKEN,
                  operations=operations, 
                  create_pr=True)

If your dataset is large (> 500 GB), you can upload it in chunks by adding the multi_commits=True, multi_commits_verbose=True, argument. More options are available in the official documentation.

FAQ

1. How to upload a large file > 50 GB?

Note that HuggingFace limits the size of each file to 50 GB. If your dataset is larger, you can split it into smaller files:

zip -s 50g input.zip --out output.zip

This command will split the input.zip into multiple files of 50GB size output.zip, output.z01, output.z01, ... You can then extract them using one of the following commands:

# option 1: combine 
zip -s0 output.zip --out input.zip

# option 2: using 7z to unzip directly
7z x output.zip

2. How to increase download speed?

If you are running on a machine with high bandwidth, you can increase your download speed by adding the following environment variable:

pip install huggingface_hub[hf_transfer]
export HF_HUB_ENABLE_HF_TRANSFER=1

Publications

• BOP Challenge 2023 on Detection, Segmentation and Pose Estimation of Seen and Unseen Rigid Objects

  • T. Hodaň, M. Sundermeyer, Y. Labbé, V. N. Nguyen, G. Wang, E. Brachmann, B. Drost, V. Lepetit, C. Rother, J. Matas
  • IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW, CV4MR workshop) 2024, Seattle
  • PDF, SLIDES, VIDEO, BIB

• BOP Challenge 2022 on Detection, Segmentation and Pose Estimation of Specific Rigid Objects

  • M. Sundermeyer, T. Hodaň, Y. Labbé, G. Wang, E. Brachmann, B. Drost, C. Rother, J. Matas
  • IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW, CV4MR workshop) 2023, Vancouver
  • PDF, SLIDES, VIDEO 1, VIDEO 2, BIB

• BOP Challenge 2020 on 6D Object Localization

  • T. Hodaň, M. Sundermeyer, B. Drost, Y. Labbé, E. Brachmann, F. Michel, C. Rother, J. Matas
  • European Conference on Computer Vision Workshops (ECCVW) 2020, Glasgow
  • PDF, SLIDES, BIB

• BOP: Benchmark for 6D Object Pose Estimation

  • T. Hodaň, F. Michel, E. Brachmann, W. Kehl, A. G. Buch, D. Kraft, B. Drost, J. Vidal, S. Ihrke, X. Zabulis, C. Sahin, F. Manhardt, F. Tombari, T.-K. Kim, J. Matas, C. Rother
  • European Conference on Computer Vision (ECCV) 2018, Munich
  • PDF, SLIDES, POSTER, BIB

The online evaluation system has been developed by T. Hodaň and A. Melenovský.