Temporal Feature Fusion for 3D Detection in Monocular Video

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Monocular 3D detection typically relies on single-frame analysis.
• Temporal and motion information in videos is valuable but under-explored in monocular 3D detection.

Purpose of the Study:

• To propose an effective and efficient temporal feature fusion method for monocular 3D detection.
• To improve 3D detection accuracy by incorporating temporal information from video sequences.

Main Methods:

• Utilizing optical flow to transform and fuse features from prior frames into the current frame.
• Introducing a scene feature propagation mechanism to accumulate historical scene information efficiently.
• Employing forward-backward scene consistency to remove occluded areas.

Main Results:

• The proposed method significantly improves performance across various monocular 3D detection baselines.
• Demonstrated model-agnostic applicability and excellent transferability.
• Achieved substantial gains in 3D detection accuracy by integrating temporal features.

Conclusions:

• The developed temporal feature fusion method is effective and computationally efficient for monocular 3D detection.
• Scene feature propagation offers a practical way to leverage historical data, mitigating computational overhead.
• This approach enhances 3D reasoning capabilities in monocular video analysis.