Learning Invariant Inter-pixel Correlations for Superpixel Generation

Authors: Sen Xu, Shikui Wei, Tao Ruan, Lixin Liao

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The experimental results on four benchmark datasets demonstrate the superiority of our approach to existing state-of-theart methods, regarding boundary adherence, generalization, and efficiency.
Researcher Affiliation Collaboration Sen Xu1,2, Shikui Wei*1,2, Tao Ruan3, Lixin Liao4 1Institute of Information Science, Beijing Jiaotong University 2Beijing Key Laboratory of Advanced Information Science and Network Technology 3Frontiers Science Center for Smart High-speed Railway System, Beijing Jiaotong University 4 Dao AI Robotics Inc.
Pseudocode Yes Algorithm 1: Training pseudocode for CDS.
Open Source Code Yes Code and pre-trained model are available at https://github.com/rookiie/CDSpixel.
Open Datasets Yes We evaluate our method on four segmentation datasets from different domains: BSDS500 (Arbelaez et al. 2010), NYUv2 (Silberman et al. 2012), KITTI (Geiger, Lenz, and Urtasun 2012), Pascal VOC2012 (Everingham et al. 2015).
Dataset Splits No The paper states: "we only train our model on the BSDS500 dataset and run inference on the other datasets." It mentions a "reduced validated set used by (Gadde et al. 2016) of Pascal VOC2012" in the context of a downstream task, but this is not a general validation split for their own model's training on BSDS500, nor are specific train/validation/test splits provided for BSDS500.
Hardware Specification Yes We use the gradient map as the auxiliary modality and conduct all the experiments on single RTX3090 GPU.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies.
Experiment Setup Yes During the training phase, we apply data augmentation through random resize, random cropping to 208 208, and random horizontal/vertical flipping for our CDS. We trained the models using Adam optimizer. The learning rate starts at 5e-4 and is updated by the poly learning rate policy. We trained our model for 150k iterations with batch size eight, and superpixel grid size d is set to 16.