Probabilistic Orientation Estimation with Matrix Fisher Distributions

Authors: David Mohlin, Josephine Sullivan, Gérald Bianchi

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We test our proposed approach on three separate datasets Pascal3D+, Model Net10-SO(3) and UPNA head pose. Table 1 compares the performance of our method and recent high performing approaches.
Researcher Affiliation Collaboration David Mohlin KTH/Tobii Stockholm, Sweden davmo@kth.se Gérald Bianchi Tobii Danderyd, Sweden gerald.bianchi@tobii.com Josephine Sullivan KTH Stockholm, Sweden sullivan@kth.se
Pseudocode No The paper describes the approach mathematically and textually but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/Davmo049/Public_prob_orientation_estimation_with_matrix _fisher_distributions
Open Datasets Yes Pascal3D+ [27] has 12 rigid object classes and contains images from Pascal VOC and Image Net of these classes. Model Net10-SO(3)[12] is a synthetic dataset. UPNA head pose [1] consists of videos with synchronized annotations of keypoints for the face in the image as well as its 3D rotation and position.
Dataset Splits Yes We have done all development and hyper-parameter optimization where the full training set was partitioned into a training and validation set. For the Pascal3D+ dataset, we use the Image Net validation split for the test set.
Hardware Specification No The paper does not provide specific details about the hardware used for experiments, such as GPU models, CPU specifications, or memory.
Software Dependencies No The paper mentions 'pytorch s implementation of SVD' but does not provide specific version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes We fine-tune the embedding and fully connected layer weights for 2 epochs. We use SGD and start with a learning rate of 0.01. We use a batch size of 32 and train for 120 epochs. For Pascal3D+ we reduce this learning rate by a factor 10 at epochs 30, 60 and 90. For Model Net10-SO(3) we train for 50 epochs and reduce the learning rate by a factor of 10 at epochs 30, 40 and 45. For UPNA head pose we use the same hyperparameters as for Pascal3D+, except we do not use a class embedding since there are only faces in this dataset.