Multi-Scale Context Aggregation by Dilated Convolutions

Authors: Fisher Yu, Vladlen Koltun

ICLR 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Using the simplified prediction module, we evaluate the presented context network through controlled experiments on the Pascal VOC 2012 dataset (Everingham et al., 2010). The experiments demonstrate that plugging the context module into existing semantic segmentation architectures reliably increases their accuracy.
Researcher Affiliation Collaboration Fisher Yu Princeton University Vladlen Koltun Intel Labs
Pseudocode No The paper describes the network architecture in text and Table 1, but it does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes The trained models can be found at https://github.com/fyu/dilation.
Open Datasets Yes Using the simplified prediction module, we evaluate the presented context network through controlled experiments on the Pascal VOC 2012 dataset (Everingham et al., 2010). ... For fair comparison with recent high-performing systems, we trained a front-end module that has the same structure as described in Section 4, but is trained on additional images from the Microsoft COCO dataset (Lin et al., 2014).
Dataset Splits Yes This simplified prediction module was trained on the Pascal VOC 2012 training set, augmented by the annotations created by Hariharan et al. (2011). We did not use images from the VOC-2012 validation set for training and therefore only used a subset of the annotations of Hariharan et al. (2011). ... We use the split of Sturgess et al. (2009), which partitions the dataset into 367 training images, 100 validation images, and 233 test images.
Hardware Specification No No specific hardware (e.g., GPU model, CPU type, memory) used for experiments is mentioned in the paper.
Software Dependencies No Our implementation is based on the Caffe library (Jia et al., 2014). Our implementation of dilated convolutions is now part of the stanfard Caffe distribution.
Experiment Setup Yes Training was performed by stochastic gradient descent (SGD) with mini-batch size 14, learning rate 10^-3, and momentum 0.9. The network was trained for 60K iterations.