Discriminative Deep Hashing for Scalable Face Image Retrieval
Authors: Jie Lin, Zechao Li, Jinhui Tang
IJCAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on two datasets demonstrate that the proposed method achieves superior performance compared with some state-of-the-art hashing methods. |
| Researcher Affiliation | Academia | School of Computer Science and Engineering, Nanjing University of Science and Technology jinhuitang@njust.edu.cn |
| Pseudocode | No | No pseudocode or algorithm block was found in the paper. |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described. |
| Open Datasets | Yes | You Tube Faces. This dataset contains 3,425 videos involved with 1,595 different persons... We randomly selected 40 face images per person as the training set and 5 images per person as the test set. Thus, we have 63800 images as the training set and 7975 images as the test data. All face images are resized to 32 × 32. Face Scrub. It composes a total of 106,863 face images of 530 celebrities, with about 200 images per person. So far, it is one of the largest public face datasets. We randomly select 5 face images per person as the test data, and the remainder as the training set. All face images are also resized to 32 × 32. |
| Dataset Splits | Yes | We use a training set of 10 percent as a validation set to verify the results of the training set. |
| Hardware Specification | Yes | All the experiments are implemented with Keras framework on a NVIDIA GTX 850M with CUDA7.5 and cu DNN v5.1. |
| Software Dependencies | Yes | All the experiments are implemented with Keras framework on a NVIDIA GTX 850M with CUDA7.5 and cu DNN v5.1. |
| Experiment Setup | Yes | The filter size in first convolutional layer is 3 × 3 with stride 1, and the other convolutional layers employ the filter with the size of 2 × 2 with stride 1. Weights in the fourth convolutional layer are totally unshared within the region of 2 × 2. The numbers of feature maps of the four convolutional layers are set to 20, 40, 60 and 80 respectively. For the activation function, all convolutional layers and the Face Feature layer use Rectification Linear Unit (ReLU) [Krizhevsky et al., 2012]. For the parameter setting of the divide-and-encode module, the number of each group on the Slice layer is set to 4 in our experiments. In addition, we apply Batch Normalization [Ioffe and Szegedy, 2015] after each convolutional layer to accelerate the convergence speed of the objective function. Furthermore, we adopt Adam [Kingma and Ba, 2014] as our optimization algorithm and fix the batch size as 256 during training. We use a training set of 10 percent as a validation set to verify the results of the training set. For the parameters α and β, in experiments we set them to 0.0002 and 1, respectively. |