Weakly-Supervised Contrastive Learning for Unsupervised Object Discovery.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised object discovery (UOD) aims to identify objects without labeled data, crucial for localization and segmentation.
• Existing UOD methods include generative approaches and clustering based on self-supervised models.
• Generative methods depend on reconstruction quality, while clustering methods struggle with semantic correlations.

Purpose of the Study:

• To propose a novel approach for unsupervised object discovery.
• To enhance semantic information exploration in self-supervised models for UOD.
• To improve the generic object discovery capabilities without labeled datasets.

Main Methods:

• A semantic-guided self-supervised learning model was designed, fine-tuning the DINO model's encoder via weakly-supervised contrastive learning (WCL).
• Principal Component Analysis (PCA) was employed to localize object regions based on extracted semantic features.
• The principal projection direction with the maximal eigenvalue was used as an object indicator.

Main Results:

• The proposed method effectively enhances semantic information exploration for object discovery.
• Experiments on benchmark datasets demonstrate the effectiveness of the WCL-enhanced approach.
• The integration of WCL with DINO and PCA shows promising results in UOD.

Conclusions:

• The novel approach successfully addresses limitations in existing unsupervised object discovery techniques.
• Weakly-supervised contrastive learning significantly improves semantic feature representation for UOD.
• The method offers a robust solution for generic object discovery and localization.