A Highly Efficient Model to Study the Semantics of Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Network (CNN)-based salient object detection (SOD) methods show high performance.
• However, the encoding of semantic information and category-agnostic nature of SOD models remain underexplored.
• ImageNet pre-trained backbones in SOD models can lead to information leakage and feature redundancy.

Purpose of the Study:

• To investigate the semantic encoding and category-agnostic properties of SOD models.
• To develop a lightweight, task-specific SOD model independent of classification backbones.
• To analyze the necessity of ImageNet pre-training for SOD.

Main Methods:

• Proposed CSNet, an extremely lightweight holistic model for SOD, trained from scratch.
• Employed a novel dynamic weight decay scheme for representation redundancy reduction.
• Evaluated performance on popular SOD benchmarks against state-of-the-art (SOTA) models.

Main Results:

• CSNet, with only 100K parameters (0.2% of large models), achieves performance on par with SOTA.
• Demonstrated that SOD and classification methods utilize different underlying mechanisms.
• Confirmed that SOD models are category-insensitive.
• Showed that ImageNet pre-training is not essential for effective SOD training.

Conclusions:

• SOD models operate differently from classification models and are inherently category-insensitive.
• Lightweight, task-specific models like CSNet are sufficient and efficient for SOD.
• Eliminating reliance on ImageNet pre-training reduces model complexity and redundancy in SOD.