An Analysis of SVD for Deep Rotation Estimation

Authors: Jake Levinson, Carlos Esteves, Kefan Chen, Noah Snavely, Angjoo Kanazawa, Afshin Rostamizadeh, Ameesh Makadia

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive quantitative analysis shows simply replacing existing representations with the SVD orthogonalization procedure obtains state of the art performance in many deep learning applications covering both supervised and unsupervised training. and Our contributions include A theoretically motivated analysis of rotation estimation via SVD orthogonalization in the context of neural networks... An extensive quantitative evaluation of SVD orthogonalization spanning four diverse application environments...
Researcher Affiliation Collaboration Jake Levinson1 Carlos Esteves2 Kefan Chen3 Noah Snavely3 Angjoo Kanazawa3 Afshin Rostamizadeh3 Ameesh Makadia3 1Simon Fraser University 2University of Pennsylvania 3Google Research
Pseudocode No No structured pseudocode or algorithm blocks were found.
Open Source Code No The paper does not provide a concrete link to source code or explicitly state that the source code for their methodology is publicly available.
Open Datasets Yes Images are rendered from Model Net10 [51] objects from arbitrary viewpoints. and Pascal3D+ [52] is a standard benchmark for object pose estimation from single images. and unsupervised learning of depth and ego-motion from videos [55]. ... on KITTI [8]
Dataset Splits No The paper does not provide specific details on training, validation, and test dataset splits with percentages, sample counts, or explicit splitting methodology, but refers to standard benchmarks.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory used for running its experiments.
Software Dependencies No The procedure can easily be used in popular deep learning libraries (e.g. Py Torch [34] and Tensor Flow [1] both provide differentiable SVD ops).
Experiment Setup Yes The model architecture follows the architecture described in [57]. Point clouds are embedded with simplified Point Net (4-layer MLP) ending with a global max-pooling. Three dense layers make up the regression network. and For SVD, 6D, 5D, QCQP, and the classic representations, the loss is L(R, Rt) = 1/2 ||R - Rt||^2_F. and All models are trained for 550K steps in this case.