Spatial Mixture Models with Learnable Deep Priors for Perceptual Grouping

Authors: Jinyang Yuan, Bin Li, Xiangyang Xue9135-9142

AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive empirical tests on two perceptual grouping datasets demonstrate that the proposed method outperforms the stateof-the-art methods under most experimental configurations. We evaluate our method on two perceptual grouping datasets, in which images are composed of simple shapes or handwritten images, under different experimental configurations. Extensive empirical results suggest that representing the complex regions of background pixels in a compositional manner is crucial to high-quality grouping results.
Researcher Affiliation Collaboration Jinyang Yuan, Bin Li, Xiangyang Xue Shanghai Key Laboratory of Intelligent Information Processing School of Computer Science, Fudan University Fudan-Qiniu Joint Laboratory for Deep Learning Shanghai Institute of Intelligent Electronics & Systems
Pseudocode Yes Algorithm 1 Proposed Method (Gaussian Distribution)
Open Source Code No The paper does not contain any statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes Our method is evaluated on two datasets derived from a set of publicly released perceptual grouping datasets provided by (Greff, Srivastava, and Schmidhuber 2015; Greff et al. 2016; Greff, van Steenkiste, and Schmidhuber 2017). We refer to these two datasets as the Multi-Shapes dataset and the Multi-MNIST dataset.
Dataset Splits No The paper mentions 'training and validation set' in the context of Multi-MNIST subset 3, but it does not provide specific percentages or counts for training, validation, or test splits across any of the datasets used.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments (e.g., GPU models, CPU types, or memory).
Software Dependencies No The paper mentions the use of 'Adam optimization algorithm' and 'convolutional neural networks' with specific layer types, but it does not specify any software libraries or their version numbers (e.g., Python, TensorFlow, PyTorch versions).
Experiment Setup Yes For Tagger, parameters of mixture models are updated via a 3-layer Ladder Network. For N-EM and the proposed method, the encoder and decoder networks are convolutional neural networks (CNNs) with 2 convolutional, 2 fully-connected and 3 layer normalization (Ba, Kiros, and Hinton 2016) layers. Neural networks are trained with the Adam optimization algorithm (Kingma and Ba 2015) for all approaches.