RankSEG: A Consistent Ranking-based Framework for Segmentation

Authors: Ben Dai, Chunlin Li

JMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this paper, we establish a theoretical foundation of segmentation with respect to the Dice/Io U metrics, including the Bayes rule and Dice-/Io U-calibration, analogous to classification-calibration or Fisher consistency in classification. ... The numerical effectiveness of Rank Dice/m Rank Dice is demonstrated in various simulated examples and Fine-annotated City Scapes, Pascal VOC and Kvasir-SEG datasets with state-of-the-art deep learning architectures.
Researcher Affiliation Academia Ben Dai EMAIL Department of Statistics The Chinese University of Hong Kong Hong Kong SAR. Chunlin Li EMAIL School of Statistics University of Minnesota MN 55455 USA.
Pseudocode Yes Algorithm 1: Computing schemes for the proposed Rank Dice framework. ... Algorithm 2: m Rank Dice for overlapping m Dice-segmentation.
Open Source Code Yes Python module and source code are available on GITHUB at https://github.com/statmlben/rankseg.
Open Datasets Yes The numerical effectiveness of Rank Dice/m Rank Dice is demonstrated in various simulated examples and Fine-annotated City Scapes, Pascal VOC and Kvasir-SEG datasets with state-of-the-art deep learning architectures.
Dataset Splits Yes Pascal VOC 2012 dataset contains 1,464 training and 1,449 validation pixel-level annotated images.
Hardware Specification Yes All experiments are conducted using Py Torch and CUDA on an NVIDIA Ge Force RTX 3080 GPU.
Software Dependencies No All experiments are conducted using Py Torch and CUDA on an NVIDIA Ge Force RTX 3080 GPU. ... The experiment protocol of our numerical sections basically follows a well-developed Github repository PYTORCH-SEGMENTATION (Ouali, 2022).
Experiment Setup Yes For all methods, we employ SGD on the learning rate (lr) schedule lr schedule= poly , and the initial learning rate initial lr=0.01, weight decay=100, momentum=0.9, crop size 512x512, batch size 6, and 300 epochs. The performance on validation set is measured in terms of the m Dice and m Io U averaged across 19 object classes (Table 2).