IDNet: Information Decomposition Network for Fast Panoptic Segmentation.

Area of Science:

• Computer Vision
• Deep Learning

Background:

• Traditional panoptic segmentation methods use separate instance and semantic segmentation branches, leading to complex and time-consuming pipelines.
• These methods involve excessive feature extraction and result fusion, hindering efficiency.

Purpose of the Study:

• To propose IDNet, a novel approach for panoptic segmentation that decomposes the task at an information level.
• To simplify and unify the panoptic segmentation process for both 'stuff' and 'things' categories.

Main Methods:

• IDNet decomposes panoptic segmentation into category and location information extraction.
• A serial pipeline recomposes these information types, streamlining the process.
• Two specialized correction losses are employed to ensure prediction accuracy.

Main Results:

• IDNet achieves the fastest inference speed of 24.2 FPS at 800x1333 resolution on a Tesla V100 GPU.
• The model attains a Panoptic Quality (PQ) score of 43.8, comparable to one-stage CNN-based methods.
• IDNet demonstrates a superior balance between effectiveness and computational efficiency.

Conclusions:

• IDNet offers a significantly simplified and unified approach to panoptic segmentation.
• The method achieves state-of-the-art efficiency without compromising performance.
• The proposed information-level decomposition and serial pipeline represent a promising direction for future research in panoptic segmentation.