Multi-Scale Bidirectional FCN for Object Skeleton Extraction
Authors: Fan Yang, Xin Li, Hong Cheng, Yuxiao Guo, Leiting Chen, Jianping Li
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on various commonly used benchmarks demonstrate that the proposed MSBFCN has achieved significant improvements over the state-ofthe-art algorithms. |
| Researcher Affiliation | Academia | 1School of Computer Science & Engineering, University of Electronic Science and Technology of China 2Center for Robotics, School of Automation Engineering, University of Electronic Science and Technology of China |
| Pseudocode | No | The paper describes methods in text and mathematical formulas but does not provide pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not explicitly state that the source code for the described methodology is publicly available, nor does it provide a link to a code repository. |
| Open Datasets | Yes | We evaluate the proposed MSB-FCN model on four widely-used benchmark datasets: SK-SMALL (Shen et al. 2016c), WH-SYMMAX (Shen et al. 2016b), SKLARGE (Shen et al. 2016a) and Sym-PASCAL (Ke et al. 2017). |
| Dataset Splits | No | The paper specifies training and testing splits for each dataset (e.g., 'SK-SMALL... includes 300 training images and 206 testing images.'), but does not explicitly mention a separate validation set or its split. |
| Hardware Specification | Yes | The traditional methods are tested on a PC with an i7 2.50 GHz CPU and 8 GB RAM, while the CNN-based methods are accelerated by a NVIDIA GTX 1080ti GPU X 11G. |
| Software Dependencies | No | The paper states 'our network is based on the publicly available platform Caffe (Jia et al. 2014),' but does not specify a version number for Caffe or any other key software dependencies. |
| Experiment Setup | Yes | The input image is resized such that its resolutions become 480 480 pixels. In Eq. 12, we define ξsf = 2 and γsm = 1 so as to emphasize the final output. We use the poly learning rate policy (Liu, Rabinovich, and Berg 2015), where the learning rate is automatically controlled by (1 iter maxiter )power. The initial learning rate is set to 10 8, and the power is set to 0.9. We set the maximum number of iterations to 60K. The Stochastic Gradient Descent (SGD) is employed for optimization. The outputs of different scales are also resized to 480 480 pixels to compute the loss. To reduce overfitting, the training data is augmented by rotating all the training images by every 90 degrees, flipping them with different axes, and resizing them to three different scales (0.8, 1.0, 1.2), following (Shen et al. 2016c). |