LabelMaker

Installation

This is an example on Ubuntu 20.02 with cuda 11.8.

Environment for LabelMaker

This environment is used for semantic segmentation of several models, and it is also used for generating consensus semantic labels.

bash env_v2/install_labelmaker_env.sh 3.9 11.3 1.12.0 9.5.0

This command creates a conda environment called labelmaker with python version 3.9, cuda version 11.8, pytorch version 2.0.0, and gcc version 10.4.0. Here are possible sets of environment versions:

Python	CUDA toolkit	PyTorch	GCC
3.9	11.3	1.12.0	9.5.0
3.9	11.6	1.13.0	10.4.0
3.9	11.8	2.0.0	10.4.0
3.10	11.8	2.0.0	10.4.0

For python=3.10, I only tested with 3.10 11.8 2.0.0 10.4.0, others might also be possible.

conda activate labelmaker

Environment for SDFStudio

This environment is used for generating consistent consensus semantic labels. It use the previous consensus semantic labels (together with RGBD data) to train a neural implicit surface and get a view-consistent consensus semantic label. It uses a modified version of SDFStudio. SDFStudio need specific version of pytorch, therefore, it is made as a separate environment. To install the environment, run

bash env_v2/install_sdfstudio_env.sh 3.10 11.3

Python=3.10 and CUDA-toolkit==11.3 is the only tested combination. This version of SDFStudio requires torch==1.12.1, which only supports CUDA 11.3 and 11.6, therefore, it might be impossible to run it on newer GPUs.

conda activate sdfstudio

Download Model Checkpoints

bash env_v2/download_checkpoints.sh

Docker Image

Docker image based on Ubuntu 16.04

# Build
docker build --tag labelmaker-env-16.04 -f docker/ubuntu16.04+miniconda.dockerfile .

# Run
docker run \
  --gpus all \
  -i --rm \
  -v ./env_v2:/LabelMaker/env_v2 \
  -v ./models:/LabelMaker/models \
  -v ./labelmaker:/LabelMaker/labelmaker \
  -v ./checkpoints:/LabelMaker/checkpoints \
  -v ./testing:/LabelMaker/testing \
  -v ./.gitmodules:/LabelMaker/.gitmodules \
  -t labelmaker-env-16.04 /bin/bash

Docker image based on Ubuntu 20.04

# Build
docker build --tag labelmaker-env-20.04 -f docker/ubuntu20.04+miniconda.dockerfile .

# Run
docker run \
  --gpus all \
  -i --rm \
  -v ./env_v2:/LabelMaker/env_v2 \
  -v ./models:/LabelMaker/models \
  -v ./labelmaker:/LabelMaker/labelmaker \
  -v ./checkpoints:/LabelMaker/checkpoints \
  -v ./testing:/LabelMaker/testing \
  -v ./.gitmodules:/LabelMaker/.gitmodules \
  -t labelmaker-env-20.04 /bin/bash

Setup Scene

Download scene

export TRAINING_OR_VALIDATION=Training
export SCENE_ID=47333462
python 3rdparty/ARKitScenes/download_data.py raw --split $TRAINING_OR_VALIDATION --video_id $SCENE_ID --download_dir /tmp/ARKitScenes/ --raw_dataset_assets lowres_depth confidence lowres_wide.traj lowres_wide lowres_wide_intrinsics vga_wide vga_wide_intrinsics

Convert scene to LabelMaker workspace

WORKSPACE_DIR=/home/weders/scratch/scratch/LabelMaker/arkitscenes/$SCENE_ID
python scripts/arkitscenes2labelmaker.py --scan_dir /tmp/ARKitScenes/raw/$TRAINING_OR_VALIDATION/$SCENE_ID --target_dir $WORKSPACE_DIR

Run Pipeline on Scene

Run individual models

InternImage

python models/internimage.py --workspace $WORKSPACE_DIR

OVSeg

python models/ovseg.py --workspace $WORKSPACE_DIR

Grounded SAM

python models/grounded_sam.py --workspace $WORKSPACE_DIR

python models/omnidata_depth.py --workspace $WORKSPACE_DIR
python models/hha_depth.py --workspace $WORKSPACE_DIR
python models/cmx.py --workspace $WORKSPACE_DIR

Mask3D

python models/mask3d_inst.py --workspace $WORKSPACE_DIR

OmniData normal (used for NeuS)

python models/omnidata_normal.py --workspace $WORKSPACE_DIR

Run consensus voting

python labelmaker/consensus.py --workspace $WORKSPACE_DIR

Run 3D Lifting

Point-based lifting

python -m labelmaker.lifting_3d.lifting_points --workspace $WORKSPACE_DIR

NeRF-based lifting (required for dense 2D labels)

bash labelmaker/lifting_3d/lifting.sh $WORKSPACE_DIR

Visualization

Visualize 3D point labels (after running point-based lifting)

 python -m labelmaker.visualization_3d --workspace $WORKSPACE_DIR

Bibtex

When using LabelMaker in acamdemic works, please use the following reference:

@inproceedings{Weder2024labelmaker,
  title = {{LabelMaker: Automatic Semantic Label Generation from RGB-D Trajectories}},
  author={Weder, Silvan and Blum, Hermann and Engelmann, Francis and Pollefeys, Marc},
  booktitle = {International Conference on 3D Vision (3DV)},
  year = {2024}
}

License

LabelMaker itself is released under BSD-3-clause License. However, inidividual models that can be used as part of LabelMaker may have more restrictive licenses. If a user is prohibited by license to use a specific model they can just leave them out of the pipeline. Here are the models and the licenses they use:

ARKitScenes: CC BY-NC-SA 4.0 license
InternImage: MIT
Mask3D: MIT
GSAM: Apache-2.0
OpenAI CLIP: MIT
Grounding DINO: Apache-2.0
Omnidata: custom license, view
CMX: MIT
OVSeg: Attribution-NonCommercial 4.0 International, view

cvg/LabelMaker