Low-Shot Video Object Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Video object segmentation (VOS) typically requires extensive pixel-level annotations, which are costly and time-consuming to acquire.
• Existing VOS methods are heavily reliant on dense annotations for effective training.

Purpose of the Study:

• To demonstrate the feasibility of training VOS models with sparse annotations, specifically one or two labeled frames per video.
• To introduce a novel training methodology named low-shot video object segmentation (low-shot VOS).

Main Methods:

• Developed a method for generating reliable pseudo-labels for unlabeled frames during training.
• Integrated pseudo-labeling with sparsely labeled frames to optimize VOS models.
• Proposed a universal approach for training VOS models on one-shot and two-shot datasets.

Main Results:

• Achieved performance comparable to fully annotated models using only 7.3% (YouTube-VOS) and 2.9% (DAVIS) of labeled data in a two-shot setting.
• Demonstrated a universal method for one-shot and two-shot VOS training.
• Observed a minor performance decrease in the one-shot setting compared to fully annotated training.

Conclusions:

• Low-shot VOS is a viable and efficient alternative to traditional VOS training methods.
• The proposed pseudo-labeling strategy significantly reduces the need for extensive annotations while maintaining high performance.
• This methodology offers a practical solution for VOS tasks where data annotation is a bottleneck.