VISMA-tracker

This is a preemptive release of the code for our ECCV 18 paper:

@inproceedings{feiS18,
    title = {Visual-Inertial Object Detection and Mapping},
    author = {Fei, X. and Soatto, S.},
    booktitle = {Proceedings of the European Conference on Computer Vision},
    year = {2018}
}

The data utilities released earlier can be found here.

The problem we want to address here is object detection and 6 DoF (Degrees-of-Freedom) object pose estimation.

The code provides a fusion framework (written in C++) to fuse likelihood scores from semantic modules (e.g. object detectors) and low-level image cues (e.g. edges/intensity values) to accompalish this.

For object likelihoods, the system relies on external modules such as Faster R-CNN running in TensorFlow or Pytorch. Since lots of popular deep learning models are written in Python, we provide a message-based inter-process communication facility, enabled by ZMQ (ZeroMQ) library.

For low-level image cues, the code contains implementation of various model-based tracking algorithms which leverage edges/intensity values as evidences and use gradient-based optimization/particle filtering as the underlying inference machinery.

Applications

In the app folder under the project root directory, we provide several applications using our library.

• SORBT_XXX: Single-Object Region-Based Tracker for dataset XXX.
• SODFT_XXX: Single-Object Distance-Field based Tracker for dataset XXX.
• linemod, rigidpose are two model-based tracking datasets.
• visma is our own dataset available here.

Build

We include some dependencies in the thirdparty folder. Other dependencies (listed below) should be availabe as debian packages.

To build, simply trigger build.sh in the project root directory. Missing packages should be easily resolved by looking up and installing proper packages via your favoriate package manager.

Dependencies

Numeric

• GMP: Gnu Multi-Precision
• GLM: OpenGL Mathematics
• Eigen: Template linear algebra library

Utilities

• tbb: Threading Building Blocks for CPU parallelism from Intel
• glog: logging
• googletest: unit testing
• gflags: command-line options
• jsoncpp: json for configuration

Graphics and geometry processing

• igl: Mesh loading and processing
• OpenGL: rendering

Messaging

• ZMQ: Zero Message Queue
• LCM: Lgihtweight Communications and Marshalling

Launch faster-rcnn for object likelihood

TODO: will release our customized Detectron software which provides the communication functionality.

To run the following likelihood evaluation process before launching the tracker in detectron root directory with edge branch:

python2 vlslam_module/infer_process.py --cfg configs/12_2017_baselines/fast_rcnn_R-50-FPN_2x.yaml --output-dir /tmp/detectron-visualizations --wts models/faster_rcnn_R-50-FPN_2x.pkl