Learning FRAME Models Using CNN Filters

Authors: Yang Lu, Song-Chun Zhu, Ying Wu

AAAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 7 Image generation experiments In our experiments, we use VGG filters (Simonyan and Zisserman 2015), and we use the Matlab code of Mat Conv Net (Vedaldi and Lenc 2014). Experiment 1: generating object patterns. We learn the non-stationary FRAME model (5) from images of aligned objects. The images are collected from the internet.
Researcher Affiliation Academia Yang Lu, Song-Chun Zhu, Ying Nian Wu Department of Statistics, University of California, Los Angeles, USA
Pseudocode Yes Algorithm 1 Learning and sampling algorithm
Open Source Code Yes The code, data, and more experimental results can be found at http://www.stat.ucla.edu/ yang.lu/project/deepFrame/main.html
Open Datasets Yes The code, data, and more experimental results can be found at http://www.stat.ucla.edu/ yang.lu/project/deepFrame/main.html
Dataset Splits No No specific details on train/validation/test splits are provided in the paper.
Hardware Specification No The paper does not provide specific hardware details (like CPU/GPU models) used for running the experiments.
Software Dependencies No The paper mentions 'Matlab code of Mat Conv Net', but does not provide specific version numbers for these software components.
Experiment Setup Yes For each category, the number of training images is around 10. We use M = 16 parallel chains for Langevin sampling. The number of Langevin iterations between every two consecutive updates of the parameters is L = 100. In the first scenery experiment, we learn 10 filters at the 4th convolutional layer (without local max pooling), based on the pre-trained VGG filters at the 3rd convolutional layer. The size of each Conv4 filter to be learned is 11 11 256. In the sunflower and egret experiments, we learn 20 filters of size 7 7 256 (with local max pooling).