WatchPose: A View-Aware Approach for Camera Pose Data Collection in Industrial Environments.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Collecting correlated scene images and camera poses is crucial for training absolute camera pose regression models.
• Data acquisition in industrial environments is challenging due to varied object sizes and non-constant motion, leading to model sensitivity.

Purpose of the Study:

• To present WatchPose, an efficient method for collecting camera pose data to enhance the generalization and robustness of camera pose regression models.
• To address the challenges of data collection in constricted industrial environments.

Main Methods:

• WatchPose utilizes nested marker tracking and augmented reality (AR) visualization to guide users.
• It facilitates the collection of training data from diverse viewpoints and broader camera-object distances.

Main Results:

• Experiments demonstrate that WatchPose significantly improves the accuracy of existing camera pose regression models compared to traditional methods.
• The proposed method enhances model robustness against variations in viewpoints and distances.

Conclusions:

• WatchPose offers a simple yet effective solution for collecting high-quality camera pose data in complex industrial settings.
• The introduction of the Industrial10 dataset will foster further research in adapting camera pose regression for challenging environments.