Fully Convolutional Neural Networks with Full-Scale-Features for Semantic Segmentation

Authors: Tianxiang Pan, Bin Wang, Guiguang Ding, Jun-Hai Yong

AAAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our proposed method is evaluated on PASCAL VOC 2012, and achieves a state-of-art result. In this section, we mainly evaluate our method on VOC 2012 benchmark dataset (Everingham et al. 2010). We use the mean intersection over union (Mean Io U) results as the standard evaluation score to compare our method with others.
Researcher Affiliation Academia Tianxiang Pan, Bin Wang, Guiguang Ding, Jun-Hai Yong School of Software, Tsinghua University, Beijing 100084, China ptx9363@gmail.com, {wangbins,dinggg,yongjh}@tsinghua.edu.cn
Pseudocode No No structured pseudocode or algorithm blocks were found.
Open Source Code No Our full-scale features model is implemented in the Caffe framework and will be published soon.
Open Datasets Yes In this section, we mainly evaluate our method on VOC 2012 benchmark dataset (Everingham et al. 2010). The VOC 2012 dataset has 20 object classes with an extra background class. The original datasets are labeled with 1464 train images and 1449 validation images. (Hariharan et al. 2011) labels additional 9118 images with segmentation annotations which are commonly used for training semantic segmentation models. We use these additional annotations and test our method on both validation and test datasets.
Dataset Splits Yes The original datasets are labeled with 1464 train images and 1449 validation images.
Hardware Specification Yes Fine-tuning the model on a NVIDIA Ge Force Ti Tan GPU costs about 1 to 1.5 days.
Software Dependencies No Our full-scale features model is implemented in the Caffe framework and will be published soon. No specific version number for Caffe or other software dependencies is provided.
Experiment Setup Yes SGD with mini-batch is used for training. A training minibatch of 10 images is taken and the initial learning rate is 1e-9 with a step learning policy that multiplied by 0.1 each 20k iterations. We use the momentum of 0.9 and the weight decay of 0.0005.