LabelFusion

This repo holds the code for LabelFusion.

• paper: https://ieeexplore.ieee.org/abstract/document/8460950
• arxiv preprint: https://arxiv.org/abs/1707.04796
• video: https://www.youtube.com/watch?v=nc65CF1W9B0

Quick Links

• Are you looking to get started making your own data? first go get our docker image: https://hub.docker.com/r/robotlocomotion/labelfusion/
• Are you looking to download the example dataset? you can find links here to either download a 5 GB subset, or all ~500 GB of data http://labelfusion.csail.mit.edu/#data
• Are you trying to understand the organization of the data? this documents the organization of the data: https://github.com/RobotLocomotion/LabelFusion/blob/master/docs/data_organization.rst

Getting Started

First step is to download and set up our Docker.

If alternatively you'd like to natively install, this document may be helpful: https://github.com/RobotLocomotion/LabelFusion/blob/master/docs/setup.rst

Inspecting Data from LabelFusion

If you've downloaded some of LabelFusion data and would like to inspect some of it, we recommend the following:

1. Run our docker image (instructions here: https://hub.docker.com/r/robotlocomotion/labelfusion/)
2. Inside the docker image navigate to a log directory and run the alignment tool. Even though the data has already been labeled, you can inspect the results:

cd ~/labelfusion/data/logs_test/2017-06-16-20
run_alignment_tool

You should see a GUI like the following:

4. Inspect labeled images (cd path-to-labelfusion-data/logs_test/2017-06-16-20/images and browse the images)
5. Run a script to print out the overall status of the dataset (note this may take ~10-20 seconds to run for the full dataset): dataset_update_status -o

Training on Object Detection, Segmentation, and/or Pose data

LabelFusion provides training data which is able to train a variety of perception systems. This includes:

• semantic segmentation (pixelwise classification)
• 2D object detection (bounding box + classification) -- note that we provide the segmentation masks, not the bounding boxes, but the bounding boxes could be computed from the masks
• 6 DoF object poses
• 3D object detection (bounding box + classidication) -- the 3D bounding box can be computed from the 6 DoF object poses together with their mesh.
• 6 DoF camera pose - this is provided without any labeling, just through the use of the dense SLAM method we use, ElasticFusion

Please see this document to better understand how the data is structured: https://github.com/RobotLocomotion/LabelFusion/blob/master/docs/data_organization.rst

At the time of publication for LabelFusion, we used this repo to train segmentation networks: https://github.com/DrSleep/tensorflow-deeplab-resnet

Quick Pipeline Instructions for Making New Labeled Data with LabelFusion

This is the quick version. If you'd prefer to go step-by-step manually, see Pipeline_Instructions.

Collect raw data from Xtion

First, cdlf && cd data/logs, then make a new directory for your data. In one terminal, run:

openni2-camera-lcm

In another, run:

lcm-logger

Your data will be saved in current directory as lcmlog-*.

Process into labeled training data

First we will launch a log player with a slider, and a viewer. The terminal will prompt for a start and end time to trim the log, then save the outputs:

run_trim

Next, we prepare for object pose fitting, by running ElasticFusion and formatting the output:

run_prep

Next, launch the object alignment tool and follow the three steps:

run_alignment_tool

1. Check available object types:

• In your data directory, open object_data.yaml and review the available objects, and add the objects / meshes that you need.
  • If you need multiple instances of the same object, you will need to create separate copies of the object with unique names (e.g. drill-1, drill-2, ...). For networks that do object detection, ensure that you remove this distinction from your labels / classes.

2. Align the reconstructed point cloud:

   • Open measurement panel (View -> Measurement Panel), then check Enabled in measurement panel

   • Use shift + click and click two points: first on the surface of the table, then on a point above the table

   • Open Director terminal with F8 and run:

     gr.rotateReconstructionToStandardOrientation()
     

   • Close the run_alignment_tool application (ctrl + c) and rerun.

3. Segment the pointcloud above the table

   • Same as above, use shift + click and click two points: first on the

   surface of the table, then on a point above the table - Open Director terminal with F8 and run:

   gr.segmentTable()
   gr.saveAboveTablePolyData()
   

   • Close the run_alignment_tool application (ctrl + c) and rerun.

4. Align each object and crop point clouds.

   • Assign the current object you're aligning, e.g.:

     objectName = "drill"
     

   • Launch point cloud alignment:

     gr.launchObjectAlignment(objectName)
     

     This launches a new window. Click the same three points in model and on pointcloud. Using shift + click to do this. After you do this the affordance should appear in main window using the transform that was just computed.

     • If the results are inaccurate, you can rerun the above command, or you can double-click on each affordance and move it with an interactive marker: left-click to translate along an axis, right-click to rotate along an axis.

   • When you are done with an object's registration (or just wish to save intermediate poses), run:

     gr.saveRegistrationResults()
     

After the alignment outputs have been saved, we can create the labeled data:

run_create_data

By default, only RGB images and labels will be saved. If you'd also like to save depth images, use the -d flag:

run_create_data -d

Citing LabelFusion

If you find LabelFusion useful in your work, please consider citing:

@inproceedings{marion2018label,
  title={Label Fusion: A Pipeline for Generating Ground Truth Labels for Real RGBD Data of Cluttered Scenes},
  author={Marion, Pat and Florence, Peter R and Manuelli, Lucas and Tedrake, Russ},
  booktitle={2018 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={3325--3242},
  year={2018},
  organization={IEEE}
}