Image Aesthetic Assessment Assisted by Attributes through Adversarial Learning
Authors: Bowen Pan, Shangfei Wang, Qisheng Jiang679-686
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on two benchmark databases demonstrate the superiority of the proposed method to state of the art work. |
| Researcher Affiliation | Academia | Bowen Pan,1 Shangfei Wang, ,1,2 Qisheng Jiang2 Key Lab of Computing and Communication Software of Anhui Province 1School of Computer Science and Technology, 2School of Data Science, University of Science and Technology of China, Hefei, Anhui, P.R.China, 230027 bowenpan@mail.ustc.edu.cn; sfwang@ustc.edu.cn; qishengj@mail.ustc.edu.cn |
| Pseudocode | Yes | Algorithm 1 The learning algorithm of the attributes assisted image aesthetic assessment framework. |
| Open Source Code | No | The paper does not contain any concrete access information (specific link, explicit statement of release, or mention of supplementary materials) for the source code of the described methodology. |
| Open Datasets | Yes | To the best of our knowledge, there are only two image aesthetic assessment databases containing aesthetic attributes: the Aesthetics and Attributes database (AADB) (?) and the Aesthetics Visual Analysis database (AVA) (?). |
| Dataset Splits | Yes | The official partition for the AADB database are 8,500 images for training, 500 images for validation and 1,000 images for testing. The official partition for the AVA database are 230,000 images for training, 20,000 images for testing. On the AVA database, since there is only official train/test split, 20,000 images are selected randomly from the training set as the validation set so that the validation and test set have the same size. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions 'Py Torch deep learning framework' and 'Adam algorithm' but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | Given a color image, we first rescale the image so that the shorter side is of length 256. Then, a 224 224 patch is cropped randomly from the rescaled image on the training set for the purpose of data augmentation while the central 224 224 patch is cropped on the validation/test set. The aesthetic score and numerical aesthetic attributes are normalized to the interval of [0, 1]. Binary aesthetic attributes are converted to discrete values of 0 or 1. ... For the rating network, we first extract feature representations from the pretrained Res Net-50 and the size of the feature representations is 2048D. Upon the 2048D feature representations, we build two hidden full connected layers with Re LU activations. The sizes of these two layers are 512 and 128, respectively. The last layer is the output layer with sigmoid activation since all of the aesthetic score and aesthetic attributes are in [0, 1]. The size of the output layer is determined by the specific method and database. For the discriminator, a neural network with two hidden layers is used. The size of the hidden layer is eight. We train the model using Adam algorithm (?) with a mini-batch size of 64. ... The learning rate starts from 0.001 and is divided by 10 when the performance on the validation set plateaus. |