Polar Transformer Networks

Authors: Carlos Esteves, Christine Allen-Blanchette, Xiaowei Zhou, Kostas Daniilidis

ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental PTN achieves state-of-the-art on rotated MNIST and the newly introduced SIM2MNIST dataset, an MNIST variation obtained by adding clutter and perturbing digits with translation, rotation and scaling. The ideas of PTN are extensible to 3D which we demonstrate through the Cylindrical Transformer Network. We present state-of-the-art performance on rotated MNIST and SIM2MNIST, which we introduce. EXPERIMENTS
Researcher Affiliation Academia Carlos Esteves, Christine Allen-Blanchette, Xiaowei Zhou, Kostas Daniilidis GRASP Laboratory, University of Pennsylvania {machc, allec, xiaowz, kostas}@seas.upenn.edu
Pseudocode No The paper describes the network architecture and modules in prose and diagrams, but it does not include any pseudocode or algorithm blocks.
Open Source Code Yes http://github.com/daniilidis-group//polar-transformer-networks
Open Datasets Yes Rotated MNIST (Larochelle et al., 2007) ... MNIST R, RTS are replicated from Jaderberg et al. (2015) ... We introduce SIM2MNIST ... Model Net40 (Wu et al., 2015) ... Street View House Numbers (SVHN) dataset Netzer et al. (2011)
Dataset Splits Yes Rotated MNIST... The training, validation and test sets are of sizes 10k, 2k, and 50k, respectively. ... SIM2MNIST... The training, validation and test set have size 10k, 5k, and 50k, respectively.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU models, CPU types, memory) used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes The building block of the network is a 3 × 3 K convolutional layer followed by batch normalization, an ReLU and occasional subsampling through strided convolution. ... The polar origin predictor comprises three blocks of 20 filters each, with stride 2 on the first block... S, small network, with seven blocks of 20 filters and one round of subsampling... B, big network, with 8 blocks with the following number of filters: 16, 16, 32, 32, 32, 64, 64, 64. Subsampling by strided convolution is used whenever the number of filters increase. ... Rotation and polar origin augmentation during training time, and wrap around padding all contribute to reduce the error. Results are from PTN-B on the rotated MNIST.