SWIN-RIND: Edge Detection for Reflectance, Illumination, Normal and Depth Discontinuity with Swin Transformer
LUN MIAO (The University of Tokyo),* Takeshi Oishi (The University of Tokyo), Ryoichi Ishikawa (The university of Tokyo)
The 34th British Machine Vision Conference
Abstract
Edges are caused by the discontinuities in surface-reflectance, illumination, surface-normal, and depth (RIND). However, extensive research into the detection of specific edge types has not been conducted. Thus, in this paper, we propose a Swin Transformer-based method (referred to as SWIN-RIND) to detect these four edge types from a single input image. Attention-based approaches have performed well in general edge detection and are expected to work effectively for RIND edges. The proposed method utilizes the Swin Transformer as the encoder and a top-down and bottom-up multilevel feature aggregation block as the decoder. The encoder extracts cues at different levels, and the decoder integrates these cues into shared features containing rich contextual information. Then, each specific edge type is predicted through independent decision heads. To train and evaluate the proposed model, we used the public BSDS-RIND benchmark, which is based on the Berkeley Segmentation Dataset and contains annotations for the four RIND-edge types. The proposed method was evaluated experimentally, and the results demonstrate that the proposed SWIN-RIND method outperforms several state-of-the-art methods.Video
Citation
@inproceedings{MIAO_2023_BMVC,
author = {LUN MIAO and Takeshi Oishi and Ryoichi Ishikawa},
title = {SWIN-RIND: Edge Detection for Reflectance, Illumination, Normal and Depth Discontinuity with Swin Transformer},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0707.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by