The Shape of Art History in the Eyes of the Machine
Authors: Ahmed Elgammal, Bingchen Liu, Diana Kim, Mohamed Elhoseiny, Marian Mazzone
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conducted a comprehensive study of several of the state-of-the-art convolutional neural networks applied to the task of style classification on 67K images of paintings, and analyzed the learned representation through correlation analysis with concepts derived from art history. |
| Researcher Affiliation | Academia | Ahmed Elgammal, Bingchen Liu, Diana Kim, Mohamed Elhoseiny The Art and AI Laboratory, Department of Computer Science Rutgers University, NJ, USA Marian Mazzone Department of Art History, College of Charleston, SC, USA |
| Pseudocode | No | No pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | No | No explicit statement or link providing access to the open-source code for the methodology described was found. |
| Open Datasets | Yes | Datasets: Training-testing Set: We trained, validated, and tested the networks using paintings from the publicly avail-able Wiki Art dataset1. This collection (as downloaded in 2015) has images of 81,449 paintings from 1,119 artists ranging from the fifteenth century to contemporary artists. Several prior studies on style classification used subsets of this dataset (e.g. (Karayev et al. 2013; Saleh et al. 2016; Saleh and Elgammal 2015)). For the purpose of our study we reduced the number of classes to 20 classes by merging fine-grained style classes with small number of images2. We excluded from the collections images of sculptures and photography. The total number of images used for training, validation, and testing are 76,921 images. |
| Dataset Splits | Yes | We split the data into training (85%), validation (9.5%) and test sets (5.5%). |
| Hardware Specification | No | No specific hardware (GPU, CPU models, etc.) used for running experiments was mentioned. |
| Software Dependencies | No | No specific software dependencies with version numbers were listed in the paper. |
| Experiment Setup | Yes | For all the models, the final softmax layer, originally designed for the 1000 classes in Image Net, was removed and replaced with a layer of 20 softmax nodes, one for each style class. Our study included varying the training strategies (training from scratch on art data vs. using pretrained models and fine-tuning them on art data), varying the network architecture, and data augmentation strategies. ...we added two layers with 1024 and 512 nodes to all the models, and the models are then fine-tuned to adjust the weights for the new layers. |