Decoupled Deep Neural Network for Semi-supervised Semantic Segmentation

Authors: Seunghoon Hong, Hyeonwoo Noh, Bohyung Han

NeurIPS 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that our network substantially outperforms existing semi-supervised techniques based on DNNs even with much smaller segmentation annotations, e.g., 5 or 10 strong annotations per class. Section 6 presents experimental results on a challenging benchmark dataset.
Researcher Affiliation Academia Dept. of Computer Science and Engineering, POSTECH, Pohang, Korea
Pseudocode No The paper describes the proposed methods using text and diagrams, but it does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper states 'Refer to our project website4 for more comprehensive qualitative evaluation' (Footnote 4: 'http://cvlab.postech.ac.kr/research/decouplednet/'). This is a project website for evaluation, not an explicit statement that the source code for the methodology is available, nor does it provide a specific repository link.
Open Datasets Yes We employ PASCAL VOC 2012 dataset [1] for training and testing of the proposed deep network. The dataset with extended annotations from [16], which contains 12,031 images with pixel-wise class labels, is used in our experiment. There are 10,582 images with image-level class labels, which are used to train our classification network.
Dataset Splits Yes We evaluate the performance of the proposed algorithm on 1,449 validation images. We also construct training datasets with strongly annotated images; the number of images with segmentation labels per class is controlled to evaluate the impact of supervision level.
Hardware Specification Yes training takes 3 days (0.5 day for classification network and 2.5 days for segmentation network) in a single Nvidia GTX Titan X GPU with 12G memory.
Software Dependencies No The paper states 'The proposed network is implemented based on Caffe library [17]' but does not provide a specific version number for Caffe or any other software dependency.
Experiment Setup Yes The standard Stochastic Gradient Descent (SGD) with momentum is employed for optimization, where all parameters are identical to [12]. We use mini-batches of 64 examples in training classification and segmentation networks, respectively; training takes 3 days (0.5 day for classification network and 2.5 days for segmentation network). We perform combinatorial cropping proposed in Section 5 for the images with strong annotations, where Edge Box [15] is adopted to generate region proposals and the Np(= 200) sub-images are generated for each label combination.