Building Deep Networks on Grassmann Manifolds
Authors: Zhiwu Huang, Jiqing Wu, Luc Van Gool
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The evaluations on three visual recognition tasks show that our Grassmann networks have clear advantages over existing Grassmann learning methods, and achieve results comparable with state-of-the-art approaches. |
| Researcher Affiliation | Academia | Computer Vision Lab, ETH Zurich, Switzerland VISICS, KU Leuven, Belgium {zhiwu.huang, jiqing.wu, vangool}@vision.ee.ethz.ch |
| Pseudocode | No | As most layers in the Gr Net model are expressed with complex matrix factorization functions, they cannot be simply reduced to a constructed bottom-up from element-wise calculations. In other words, the matrix backpropgation (backprop) cannot be derived by using traditional matrix that treats element-wise operations in matrix form. |
| Open Source Code | No | As the matrix factorizations are implemented well in CUDA, we will achieve the GPU version of our Gr Net for speedups. |
| Open Datasets | Yes | We utilize the popular Acted Facial Expression in Wild (AFEW) (Dhall et al. 2014) dataset. We use the HDM05 database (M uller et al. 2007) that is one of the largest-scale skeleton-based human action datasets. We employ one standard dataset named Point-and-Shoot Challenge (Pa SC) (Beveridge et al. 2013). |
| Dataset Splits | Yes | The standard protocol designed by (Dhall et al. 2014) splits the dataset into three data sets, i.e., training, validation and test data sets. |
| Hardware Specification | Yes | For training the Gr Net, we just use an i7-2600K (3.40GHz) PC without any GPUs5. |
| Software Dependencies | No | As the matrix factorizations are implemented well in CUDA, we will achieve the GPU version of our Gr Net for speedups. |
| Experiment Setup | Yes | The learning rate λ and the batch size are set to 0.01 and 30 respectively. The FRMap matrices are all initialized as random full rank matrices, and the number of them per layer is set to 16. For all the Proj Pooling layers, the number n of the instances for pooling are fixed as 4. |