Tripartite Feature Enhanced Pyramid Network for Dense Prediction.

Area of Science:

• Computer Vision and Machine Learning
• Deep Learning Architectures

Background:

• Pyramidal feature representations are crucial for dense prediction tasks requiring multi-scale visual understanding.
• Existing Feature Pyramid Networks (FPN) face limitations in feature extraction and fusion, hindering informative feature generation.

Purpose of the Study:

• To address the weaknesses of the standard Feature Pyramid Network (FPN).
• To introduce a novel Tripartite Feature Enhanced Pyramid Network (TFPN) for improved multi-scale feature learning.

Main Methods:

• Developed a feature reference module with lateral connections for adaptive bottom-up feature extraction.
• Designed a feature calibration module for spatially aligned feature fusion between adjacent layers.
• Introduced a feature feedback module to enhance the encoding capacity of the FPN architecture.

Main Results:

• The proposed TFPN consistently and significantly outperforms the vanilla FPN across evaluations.
• TFPN demonstrates superior performance in generating more powerful and informative multi-scale feature representations.
• Evaluations were conducted on object detection, instance segmentation, panoptic segmentation, and semantic segmentation tasks.

Conclusions:

• The TFPN architecture offers a substantial improvement over traditional FPN for dense prediction tasks.
• The novel modules effectively enhance feature extraction, calibration, and feedback mechanisms for richer representations.