An Object Point Set Inductive Tracker for Multi-Object Tracking and Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Multi-object tracking and segmentation (MOTS) builds upon multi-object tracking (MOT).
• High-quality embeddings are crucial for discriminative feature learning in MOTS.
• Existing methods may suffer from feature averaging in high-dimensional embeddings.

Purpose of the Study:

• To explore the relationship between segmentation and tracking in MOTS.
• To propose an efficient and effective MOTS algorithm named OPITrack.
• To improve the discriminative power of learned embeddings for better tracking performance.

Main Methods:

• Developed an efficient Object Point set Inductive Tracker (OPITrack).
• Introduced an embedding generalization training strategy to mitigate feature averaging.
• Proposed a general Trip-hard sample augmentation loss for robust embedding learning.

Main Results:

• OPITrack demonstrated promising results on two benchmark MOTS datasets.
• The proposed training strategy and loss function enhance feature discriminability.
• OPITrack achieved better performance with reduced video memory (VRAM) during training.

Conclusions:

• OPITrack effectively addresses challenges in MOTS by improving embedding quality.
• The method offers a more efficient approach to MOTS with practical benefits for training.
• This work contributes to advancing the capabilities of multi-object tracking and segmentation systems.