This is a repo for our work 《6D Matrix Fisher Distribution for Probability Pose Estimation》.
In our work, we proposed a new kind of 6D fisher distribution which could describe rotation's uncertainty on rotation manifold with 6D rotation representation, maintaining complete continuity. To construct a complete probability density function, we apply a sampling strategy to calculate the normalizing term. Besides, we choose a transformer feature extractor to get global topology information for a better distribution modeling.
Our work has just been submitted to IEEE Signal Processing Letters and is currently under review.
conda env create -f environment.yml
The dataset is available at Pascal3D+.
You could download it and unzip it at $PROJECT_DIR/datasets/
python main.py --run_name <run_name> --config_file <configs> --gpus <gpu_index>
You could supervise the training process by:
cd logs/pascal
tensorboard --log_dir=./
python scripts/visualize_pascal3d.py
Our code referred to the repo of the matrix Fisher distribution MatrixFisher and the repo of Rotation Laplace RotationLaplace.
According to the way of dataset splitting introduced in Implicit-PDF, we use ImageNet-trainval and Pascal-train images as our training data and the Pascal-val images as our testing data.