DEL: Deep Embedding Learning for Efficient Image Segmentation

Authors: Yun Liu, Peng-Tao Jiang, Vahan Petrosyan, Shi-Jie Li, Jiawang Bian, Le Zhang, Ming-Ming Cheng

IJCAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The evaluation results on BSDS500 and PASCAL Context demonstrate that our approach achieves a good tradeoff between efficiency and effectiveness. Specifically, our DEL algorithm can achieve comparable segments when compared with MCG but is much faster than it, i.e. 11.4fps vs. 0.07fps. We conduct extensive experiments on BSDS500 [Arbel aez et al., 2011] and PASCAL Context [Mottaghi et al., 2014] datasets to evaluate the proposed image segmentation algorithm.
Researcher Affiliation Academia Yun Liu1, Peng-Tao Jiang1, Vahan Petrosyan2, Shi-Jie Li1, Jiawang Bian3, Le Zhang4, Ming-Ming Cheng1 1 Nankai University 2 KTH Royal Institute of Technology 3 University of Adelaide 4 Advanced Digital Sciences Center
Pseudocode Yes The pseudo code of superpixel merging is displayed in Algorithm 1.
Open Source Code Yes The code of this paper is available at https://github.com/yun-liu/del.
Open Datasets Yes We conduct extensive experiments on BSDS500 [Arbel aez et al., 2011] and PASCAL Context [Mottaghi et al., 2014] datasets to evaluate the proposed image segmentation algorithm.
Dataset Splits Yes When training our feature embedding model on the BSDS500 dataset [Arbel aez et al., 2011] that consists of 300 trainval images and 200 test images, we augment the trainval set. When training on the PASCAL Context dataset that is divided into 7605 trainval images and 2498 test images, we only flip the trainval images for training because the number of images in this set is adequate.
Hardware Specification No The paper mentions using a 'GPU version of SLIC' but does not specify any particular GPU model (e.g., NVIDIA A100, RTX 2080 Ti), CPU, or other hardware components used for running experiments.
Software Dependencies No Our network is based on Caffe, which is a widely used deep learning framework. The paper mentions Caffe but does not specify its version or any other software dependencies with version numbers.
Experiment Setup Yes The basic learning rate is set to 1e-5. We use a weight decay of 0.0002 and batch size of 5. The learning rate policy of step is used, and we totally run SGD for 10000 iterations with step size of 8000. Each feature embedding vector vi has 64 dimensions in our design. We control each superpixel to contain about 64 pixels.