Textureless Deformable Object Tracking With Invisible Markers.

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Tracking and reconstructing deformable objects with limited texture is a significant challenge in computer vision.
• Lack of distinct features hinders accurate correspondence matching and motion tracking.

Purpose of the Study:

• To introduce a novel method using "invisible markers" for robust tracking of low-texture deformable objects.
• To generate high-quality labeled data for training machine learning algorithms in this domain.
• To present a comprehensive dataset (DOT) for benchmarking deformable object tracking and reconstruction.

Main Methods:

• Development of "invisible markers" visible only under ultraviolet (UV) light.
• Design of a novel imaging system capturing simultaneous untouched and UV-marked object views.
• Algorithm for transferring marker-based correspondences to untouched views for ground-truth labeling.

Main Results:

• Successful generation of high-quality labeled data for training learning-based algorithms.
• Creation of the DOT dataset, comprising approximately one million video frames of diverse deformable objects with 2D and 3D ground truth correspondences.
• Benchmarking revealed significant performance improvements for state-of-the-art methods when trained on the DOT dataset.

Conclusions:

• Invisible markers provide an effective solution for accurate correspondence matching and tracking of texture-deficient deformable objects.
• The DOT dataset serves as a valuable resource for advancing research in deformable object tracking and reconstruction.
• Training on the DOT dataset enhances the generalization capabilities of computer vision models on unseen data.