Content-Variant Reference Image Quality Assessment via Knowledge Distillation
Authors: Guanghao Yin, Wei Wang, Zehuan Yuan, Chuchu Han, Wei Ji, Shouqian Sun, Changhu Wang3134-3142
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Cross-dataset experiments verify that our model can outperform all NAR/NR-IQA SOTAs, even reach comparable performance with FR-IQA methods on some occasions. Our code and more detail elaborations of supplement are available: https://github.com/guanghaoyin/CVRKD-IQA. ... Experimental results show that not only our FR-teacher can produce accurate IQA scores, but also our NAR-student can significantly outperform existing NR/NAR-IQA methods, especially on the large-scale real IQA dataset. |
| Researcher Affiliation | Collaboration | Guanghao Yin1*, Wei Wang2 , Zehuan Yuan2, Chuchu Han3, Wei Ji4, Shouqian Sun1, Changhu Wang2 1 College of Computer Science and Technology, Zhejiang University, Hangzhou, China, 2 Bytedance Inc, China, 3 Huazhong University of Science and Technology, Wuhan, China, 4National University of Singapore, Singapore |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code and more detail elaborations of supplement are available: https://github.com/guanghaoyin/CVRKD-IQA. |
| Open Datasets | Yes | For IQA training datasets, we follow (Cheon et al. 2021) to choose the commonly used synthetic Kaddid10K (Lin, Hosu, and Saupe 2019), which contains 10125 LQ-FR pairs. ... The 900 training and 100 testing HQ images of DIV2K HR dataset (Agustsson and Timofte 2017) are randomly sampled at the training and testing stages of NAR-student. |
| Dataset Splits | No | The paper mentions training on Kaddid10K and evaluating on other datasets (LIVE, CSIQ, TID2013, Kon IQ-10K) which implicitly serves as testing. It also mentions 900 training and 100 testing HQ images for DIV2K. However, it does not provide explicit validation splits (e.g., percentages or counts) for the main training dataset (Kaddid10K) or the evaluation datasets, which would be necessary for full reproducibility of dataset partitioning beyond just train/test designation. |
| Hardware Specification | Yes | All the experiments were conducted on NVIDIA Tesla-V100 GPUs. ... All experiments were conducted on NVIDIA Tesla-V100 GPU. |
| Software Dependencies | No | The paper mentions using ResNet50 (He et al. 2016) and MLP-mixer (Tolstikhin et al. 2021), but it does not specify any software environments, libraries, or programming languages with version numbers (e.g., Python version, PyTorch/TensorFlow version, CUDA version). |
| Experiment Setup | Yes | Data augmentation including horizontal flip and random rotation is applied during the training. All patches are randomly cropped from the RGB image. The batch size b is set as 32. The input patch number m is set as 10 and the patch size is set as 224 224 3 to cover more local-global combined information. The number k of distilled layers in the encoder EHQ LQ is set to 18. Moreover, the initial learning rate α is 2 10 5 and the ADAM optimizer with weight decay 5 10 4 is applied. |