Towards Light-Weight and Real-Time Line Segment Detection

Authors: Geonmo Gu, Byungsoo Ko, SeoungHyun Go, Sung-Hyun Lee, Jingeun Lee, Minchul Shin726-734

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our model with two famous LSD datasets: Wireframe (Huang et al. 2018) and York Urban (Denis, Elder, and Estrada 2008). ... We train our model on Tesla V100 GPU. ... We conduct a series of ablation experiments to analyze our proposed method.
Researcher Affiliation Industry NAVER/LINE Vision {korgm403, kobiso62, powflash, shlee.mars, jglee0206, min.stellastra}@gmail.com
Pseudocode No The paper describes its methods through prose and diagrams (e.g., Figure 3 for network architecture, Figure 4 for representation) but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes Our code is available publicly 1. (Footnote 1: https://github.com/navervision/mlsd)
Open Datasets Yes We evaluate our model with two famous LSD datasets: Wireframe (Huang et al. 2018) and York Urban (Denis, Elder, and Estrada 2008).
Dataset Splits No The paper mentions training and testing on datasets ("We train our model with the training set from the Wireframe dataset and test with both Wireframe and York Urban datasets") but does not explicitly provide details about training/validation/test splits (percentages, counts, or explicit standard validation sets).
Hardware Specification Yes We train our model on Tesla V100 GPU. ... We use i Phone 12 Pro with A14 bionic chipset and Galaxy S20 Ultra with Snapdragon 865 ARM chipset.
Software Dependencies No The paper mentions using "Tensor Flow (Abadi et al. 2016) framework for model training and TFLite 2 for porting models to mobile devices" but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes Input images are resized to 320 x 320 or 512 x 512 in both training and testing... The input augmentation consists of horizontal and vertical flips, shearing, rotation, and scaling. We use Image Net (Deng et al. 2009) pre-trained weights... Our model is trained using the Adam optimizer (Kingma and Ba 2014) with a learning rate of 0.01. We use linear learning rate warm-up for 5 epochs and cosine learning rate decay... from 70 epoch to 150 epoch. We train the model for a total of 150 epochs with a batch size of 64.