Image Detection

Line Detection

Handcrafted Line Detectors:

• LSD 2008 grows line regions, fits a rectangle to the resulting set of pixels, and finally extract a line segment.
• EDLines 2011 grows the line regions in one direction only, orthogonal to the image gradient.

Learned Line Detectors :

• Wireframe : Learning to Parse Wireframes in Images of Man-Made Environments 2018.
  • ELSD : Efficient line segment detector and descriptor 2021.
  • TP-LSD: Tri-Points Based Line Segment Detector 2020, github. uses center and offset to the endpoints representation.
  • Wireframes can be further improved through a Deep hough-transform line priors 2020.
• Unsupervised :
  • L2D2: Learnable line detector and descriptor 2021, github trained from LiDAR scans.
  • SOLD² - Self-supervised Occlusion-aware Line Description and Detection 2021, github using the homography adaptation technique initially described in SuperPoint.
• Attraction Fields: a dual representation of lines through an attraction field.
  • DeepLSD: Line Segment Detection and Refinement with Deep Image Gradients, github. Decouples the distance field and line orientation field.
    • Line Segment Refinement with Optimization using vanishing points.

Pose Estimation:

• A Stable Algebraic Camera Pose Estimation for Minimal Configurations of 2D/3D Point and Line Correspondences 2018
• Structure-From-Motion Using Lines 2005, using Plucker representation, better for geometry operation. Orthonormal representation : 4-dof, better for optimization. document.

Object Detection

Nanodet 2021: NanoDet doc, NanoDet-Plus doc

• Label Assignment: Assign Guidance Module with Dynamic Soft Label Assigner. (reference : Improving Object Detection by Label Assignment Distillation 2021. use a teacher network to generate labels for the student).
• Feature Fusion: Ghost-PAN. (reference CSPPAN)