APANet: Auto-Path Aggregation for Future Instance Segmentation Prediction.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Instance segmentation models struggle with predicting future video frames due to unobservable data.
• Current methods independently process multi-level features, limiting collaborative exploitation and self-adaptation.

Purpose of the Study:

• To develop an adaptive approach for accurate future instance segmentation prediction.
• To enhance the collaborative aggregation of multi-level features for improved self-adaptation.

Main Methods:

• Proposed the Auto-Path Aggregation Network (APANet) for selective aggregation of spatio-temporal contextual information.
• Introduced an "auto-path" mechanism to connect pyramid features for hierarchical contextual information aggregation.
• Integrated APANet with Mask R-CNN and Feature Pyramid Network (FPN) for joint learning.

Main Results:

• APANet enables selective and adaptive aggregation of pyramid features tailored to different videos/frames.
• The joint learning system achieved state-of-the-art performance on three video-based instance segmentation benchmarks.
• Demonstrated improved accuracy and self-adaption in future frame prediction.

Conclusions:

• APANet effectively addresses the challenge of future instance segmentation by adaptively aggregating multi-level features.
• The proposed method offers a significant advancement in video instance segmentation prediction.
• APANet provides a robust and adaptable solution for predicting instance segmentation in unobserved future frames.