Object Affinity Learning: Towards Annotation-Free Instance Segmentation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Instance segmentation requires costly manual mask annotations.
• Existing methods using appearance cues struggle with complex scenes and background ambiguity.

Purpose of the Study:

• To develop an annotation-free instance segmentation method that overcomes limitations of appearance-based approaches.
• To leverage geometry cues (spatial continuity, motion consistency) for improved object segmentation.

Main Methods:

• Propose an affinity-based paradigm: object affinity learning.
• Object affinity learning acts as a proxy task to determine if pixels belong to the same object.
• Utilize geometry cues for feature representation learning.
• Convert learned object affinity into instance segmentation masks via graph partition algorithms.

Main Results:

• Object affinity learning significantly outperforms existing pseudo-mask-based methods.
• Demonstrated superior performance on large-scale datasets like Waymo Open Dataset and KITTI.
• Successfully addressed the challenge of annotation-free instance segmentation in complex, real-world scenarios.

Conclusions:

• Geometry cues are crucial for robust annotation-free instance segmentation.
• The proposed object affinity learning paradigm offers a more effective approach compared to appearance-based methods.
• This work reduces the dependency on manual annotations for instance segmentation tasks.