Learning Face Hallucination in the Wild
Authors: Erjin Zhou, Haoqiang Fan, Zhimin Cao, Yuning Jiang, Qi Yin
AAAI 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show our system outperforms the prior stateof-the-art methods. In this section, we present our experiments. |
| Researcher Affiliation | Collaboration | Erjin Zhou and Haoqiang Fan Tsinghua University, Beijing, China {zej11,fhq13}@mails.tsinghua.edu.cn Zhimin Cao and Yuning Jiang and Qi Yin Megvii Technology, Beijing, China {jyn,czm,yq}@megvii.com This work was done when Erjin Zhou and Haoqiang Fan were visiting students at Megvii Technology. |
| Pseudocode | No | The paper provides a 'pipeline' summary with numbered steps (1-4) in Section 3, titled 'The pipeline of Bi-channel CNN can be summarized as:'. This is a descriptive list of steps rather than a formally presented pseudocode or algorithm block. |
| Open Source Code | No | The paper does not contain any statements about making its source code publicly available, nor does it provide a link to a code repository. |
| Open Datasets | No | The paper states, 'Our system is trained from a large collection of photos crawled from the web.' and 'Our dataset contains more than 100,000 faces.' It describes the dataset's characteristics and how it was processed, but it does not provide concrete access information (e.g., a link, DOI, or specific citation for public availability) for this custom dataset. |
| Dataset Splits | Yes | The data are divided into three parts. The 60% of faces are used as training set, 20% of faces are used as validation set and the remains are left out for testing. |
| Hardware Specification | No | The paper mentions 'Benefitting from the large amount of data and recent high performance training implementation (Jia et al. 2014)', but it does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions using 'Caffe' by citing (Jia et al. 2014) as a 'high performance training implementation', but it does not specify the version number of Caffe or any other software libraries used to implement and run the experiments. |
| Experiment Setup | Yes | The size of network s input Iin is 48x48 pixels with RGB 3-channels. The network s output is HR image with 100x100 pixels and RGB 3-channels. We initialize all of the filters and fully-connected matrices from a zero-mean Gaussian distribution with standard deviation 0.001 and set all biases to 0. We train our model by stochastic gradient descent. All of the parameters are optimized by back-propagation. The data batch size is 200. The initial learning rate is 0.00001 for all layers. The update rule for parameter w Φ = {ΦF , ΦG} in k-th iteration is vk = 0.9 vk 1 + ϵ L wk 1 , wk = wk 1 vk, where v is the momentum variable, ϵ is the learning rate, and L is the objective function defined in Eq. 3. We adjust learning rate manually during training by dividing it by 10 when the validation error stops decrease. We train our model for about 5000 cycles. |