Mesh Saliency via Weakly Supervised Classification-for-Saliency CNN.

Area of Science:

• Computer Vision
• Machine Learning
• Computer Graphics

Background:

• Deep learning for mesh saliency detection faces challenges due to the difficulty of obtaining vertex-level ground truth annotations.
• Existing methods struggle with the data requirements for training accurate saliency detection models.

Purpose of the Study:

• To develop a novel weakly supervised deep learning network for mesh saliency detection that bypasses the need for extensive vertex-level annotations.
• To leverage knowledge transfer from 3D object classification to enhance mesh saliency prediction.

Main Methods:

• A novel Classification-for-Saliency CNN (CfS-CNN) was developed, trained end-to-end using only mesh class membership, not saliency ground truth.
• The network utilizes a multi-view setup and a unique two-channel structure integrating classification and saliency features.
• Knowledge transfer from 3D object classification to mesh saliency is achieved through this integrated feature approach.

Main Results:

• The CfS-CNN significantly outperforms existing state-of-the-art methods in mesh saliency detection based on extensive experimental validation.
• The proposed method demonstrates the feasibility of training deep learning models for saliency detection without requiring detailed ground truth data.
• The CfS-CNN is also applicable to scene saliency detection, with two novel applications showcased.

Conclusions:

• Weakly supervised learning offers a viable solution to the annotation bottleneck in mesh saliency detection.
• The CfS-CNN effectively transfers knowledge from classification tasks to improve saliency prediction accuracy.
• The developed network shows promise for both mesh and scene saliency analysis, opening avenues for new applications.