Multi-Path Feedback Recurrent Neural Networks for Scene Parsing

Authors: Xiaojie Jin, Yunpeng Chen, Zequn Jie, Jiashi Feng, Shuicheng Yan

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

Reproducibility Variable Result LLM Response
Research Type Experimental To verify the effectiveness of MPF-RNN, we have conducted extensive experiments over five popular and challenging scene parsing datasets, including Sift Flow (Liu, Yuen, and Torralba 2009), Barcelona (Tighe and Lazebnik 2010), Cam Vid (Brostow et al. 2008), Stanford Background (Gould, Fulton, and Koller 2009) and recently released large-scale ADE20K (Zhou et al. 2016) and demonstrated that MPFRNN is capable of greatly enhancing the discriminative power of per-pixel feature representations. We test MPF-RNN on five challenging scene parsing benchmarks, including Sift Flow (Liu, Yuen, and Torralba 2009), Barcelona (Tighe and Lazebnik 2010), Cam Vid (Brostow et al. 2008), Stanford Background (Gould, Fulton, and Koller 2009) and ADE20K. We report the quantitative results here
Researcher Affiliation Collaboration Xiaojie Jin,1 Yunpeng Chen,2 Zequn Jie,2Jiashi Feng,2 Shuicheng Yan3,2 1NUS Graduate School for Integrative Science and Engineering, NUS 2Department of ECE, NUS 3360 AI Institute
Pseudocode No Not found. The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No Not found. The paper does not provide an explicit statement about releasing its source code or a link to a code repository for the methodology described.
Open Datasets Yes Sift Flow (Liu, Yuen, and Torralba 2009), Barcelona (Tighe and Lazebnik 2010), Cam Vid (Brostow et al. 2008), Stanford Background (Gould, Fulton, and Koller 2009) and recently released large-scale ADE20K (Zhou et al. 2016)
Dataset Splits Yes ADE20K ... Containing 20K/2K/3K fully annotated scene-centric train/val/test images with 150 classes. we conduct experiments with various S on the validation set of Sift Flow. In order to fine-tune hyperparameters, we randomly extract 2K images from train set as our validation data and retrain our model using whole train set after fixing hyperparameters.
Hardware Specification Yes On a NVIDIA Titan X GPU, the training of MPF-RNN (the model in Table 3) on Sift Flow dataset finishes in about 6 hours and the testing time for an image with the resolution of 256 256 is 0.06s.
Software Dependencies No Not found. The paper mentions frameworks like Caffe and specific network architectures but does not provide specific version numbers for software dependencies.
Experiment Setup Yes The hyperparameters introduced by MPF-RNN, including S, T and λ are fine-tuned on the validation set of Sift Flow as introduced above and then fixed for other datasets where MPF-RNN uses VGG16 network. we set λ1 = λ2 = 0.3 and λ3 = 1 throughout our experiments. For models using VGG16 network, settings of hyper-parameters including learning rate, weight decay and momentum follow (Liu, Rabinovich, and Berg 2015).