Global-residual and Local-boundary Refinement Networks for Rectifying Scene Parsing Predictions
Authors: Rui Zhang, Sheng Tang, Min Lin, Jintao Li, Shuicheng Yan
IJCAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on ADE20K and Cityscapes datasets well demonstrate the effectiveness of the two refinement methods for refining scene parsing predictions. We report experimental results on two scene parsing benchmarks including ADE20K dataset [Zhou et al., 2016] and Cityscapes dataset [Cordts et al., 2016]. |
| Researcher Affiliation | Collaboration | 1 Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China, 100190. 2 Artificial Intelligence Institute, 360 company, Beijing, China, 100025. 3 University of Chinese Academy of Sciences, Beijing, China, 100039. |
| Pseudocode | No | The paper describes network architectures and configurations but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | No | The paper does not contain any explicit statement about making the source code available or provide a link to a code repository. |
| Open Datasets | Yes | We evaluate the proposed two refinement networks on two challenging scene parsing datasets, i.e. ADE20K dataset [Zhou et al., 2016] and Cityscapes dataset [Cordts et al., 2016]. The convolutional parameters of residual models [He et al., 2016] pre-trained on large scale classification datasets [Deng et al., 2009] are utilized to obtain the low-resolution predictions |
| Dataset Splits | Yes | ADE20K Dataset: ...contains 150 semantic classes for scene parsing, 20,210 images for training, 2,000 images for validation and 3,351 images for testing. Cityscape Dataset: ...There are 2,979 images in training set, 500 images in validation set and 1,525 images in test set. |
| Hardware Specification | Yes | All of our networks are trained and tested on four parallel NVIDIA Tesla K40 GPUs. |
| Software Dependencies | No | Our experiments are implemented based on MXNet platform [Chen et al., 2015b], which is efficient concerning GPU memory utilization. (Explanation: While MXNet is mentioned, no specific version number for it or any other software dependencies is provided.) |
| Experiment Setup | Yes | Standard stochastic gradient descent (SGD) with mini-batch of 4 samples is adopted for training. We use the momentum of 0.9 and weight decay of 0.0001, the same with settings during pre-training the classification model [He et al., 2016]. For training the front model, the learning rate is initialized at 0.001 for 30 epochs and then divided by 10 for another 10 epochs. After that, GRN is trained for 20 epochs in total, including 10 epochs with the learning rate of 0.001 and 10 epochs with the learning rate of 0.0001. Finally, the learning rate of 0.0001 is implemented to train LRN for 20 epochs. |