A Deep Cascade Network for Unaligned Face Attribute Classification
Authors: Hui Ding, Hao Zhou, Shaohua Zhou, Rama Chellappa
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our approach achieves significantly better performance than state-of-the-art methods on unaligned Celeb A dataset, reducing the classification error by 30.9%. The entire section 'Experiments' detailing dataset usage, training methodology, and performance evaluations confirms empirical study. |
| Researcher Affiliation | Collaboration | Hui Ding,1 Hao Zhou,2 Shaohua Kevin Zhou,3 Rama Chellappa4 1,2,4University of Maryland, College Park 3Siemens Healthineers, New Jersey |
| Pseudocode | No | The paper describes the proposed architectures and training processes using text and diagrams (Figures 1, 2, 3), but it does not include any formal pseudocode blocks or algorithm listings. |
| Open Source Code | No | The paper does not contain any explicit statements or links indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | We use the Celeb A dataset (Liu et al. 2015) in our experiments, since it has been widely used for face attributes classification. It consists of 202,599 face images collected from the Internet and annotated with 40 binary attributes. |
| Dataset Splits | Yes | As suggested in (Liu et al. 2015), 162,770 of these images are used for training, 19,867 and 19,962 are reserved for validation and testing respectively. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments (e.g., CPU/GPU models, memory specifications). It only mentions general software used like 'Caffe'. |
| Software Dependencies | No | The paper mentions 'We use Caffe (Jia et al. 2014) to implement our networks,' but it does not specify a version number for Caffe or any other software dependencies. |
| Experiment Setup | Yes | The learning rate is fixed to be 0.0001, and the network is trained for 10 epochs with batch size of 128. The FRL network is then compressed with a learning rate of 1e 7 for the hint loss training and 0.0001 for the attribute loss training. The partbased subnets are trained for 15 epochs with the weights initialized from the whole-image-based subnet. After that, the RSL and ARL are trained with a learning rate of 0.1 with all subnets fixed. Finally, a learning rate of 0.001 is applied to train the Pa W network in an end-to-end manner. Stochastic gradient descent (SGD) is used to train all the networks. The momentum and weight decay are set at 0.9 and 0.0005 for all the experiments respectively. Horizontal flipping is applied for data augmentation. |