Realworld 3D Object Recognition Using a 3D Extension of the HOG Descriptor and a Depth Camera

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• 3D object recognition is crucial for robotics and autonomous systems.
• Existing methods often struggle with real-world data variability.

Purpose of the Study:

• To develop a robust 3D object recognition approach using depth camera data.
• To address challenges in transferring models trained on synthetic data to real-world scenarios.

Main Methods:

• A 3D extension of the histogram-of-gradients descriptor was employed.
• Preprocessing techniques were applied for rotational invariance and feature dimensionality reduction.
• A classifier was trained on synthetic objects and tested on real objects captured by a depth camera.

Main Results:

• The proposed method achieved a maximum recognition accuracy of 81.5% on a real-world dataset.
• Preprocessing significantly impacted recognition accuracy and feature dimensionality.
• Challenges in adapting synthetic training data for real-world object recognition were identified.

Conclusions:

• The 3D histogram-of-gradients approach shows promise for depth-based 3D object recognition.
• Careful preprocessing is essential for bridging the synthetic-to-real domain gap.
• Further research is needed to enhance accuracy and robustness in diverse real-world conditions.