Better Dense Trajectories by Motion in Videos

Area of Science:

• Computer Vision
• Image Processing
• Video Analysis

Background:

• Current point trajectory generation methods rely on dense optical flow and consistency checks.
• These methods often miss trajectories in occluded regions, leading to artifacts around motion boundaries (MBs).

Purpose of the Study:

• To improve the accuracy of dense point trajectory generation.
• To address limitations of existing methods in handling occlusions and motion boundaries.

Main Methods:

• Exploiting motion boundaries (MBs) in videos for improved trajectory estimation.
• Estimating MBs to enhance point trajectory accuracy in discontinuous or occluded areas.
• Tracking initial feature points across frames to generate trajectories.

Main Results:

• The proposed method achieves higher accuracy in generating dense point trajectories.
• Significant improvements are observed in areas with discontinuities and occlusions.
• Experimental results show superior performance compared to state-of-the-art methods on benchmark datasets.

Conclusions:

• Focusing on motion boundaries is crucial for accurate dense point trajectory generation.
• The developed method offers a more robust solution for video analysis tasks.
• This approach enhances the reliability of trajectory data in challenging scenarios.