Bayesian Joint Modelling for Object Localisation in Weakly Labelled Images.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Traditional object localization methods often use discriminative models, localizing each object class independently.
• Weakly supervised object localization (WSOL) presents challenges due to limited annotation data.
• Existing WSOL approaches struggle with modeling inter-class relationships and background complexities.

Purpose of the Study:

• To propose a novel Bayesian joint topic modeling framework for weakly supervised object localization.
• To enhance object localization accuracy by jointly modeling multiple object classes and shared image backgrounds.
• To leverage both weakly labeled and unlabeled data for improved learning and transferability.

Main Methods:

• A generative Bayesian joint topic model is developed to capture object co-existence and enable "explaining away" inference.
• Image backgrounds are shared across object classes to better learn contextual information and improve object delineation.
• The model integrates weakly labeled and unlabeled data, facilitating the exploitation of large-scale internet image datasets.
• Bayesian formulation allows incorporation of prior knowledge and Bayesian domain adaptation for transfer learning.

Main Results:

• The proposed Bayesian joint model demonstrates superior performance in weakly supervised object localization tasks.
• Experiments on PASCAL VOC, ImageNet, and YouTube-Object videos validate the model's effectiveness.
• Joint modeling of object classes and shared backgrounds leads to improved localization accuracy and robustness.

Conclusions:

• The Bayesian joint topic modeling approach offers a significant advancement in weakly supervised object localization.
• The framework effectively handles object co-occurrence and background variations, outperforming existing methods.
• The model's ability to utilize unlabeled data and prior knowledge enhances its practical applicability and scalability.