Deep spectral improvement for unsupervised image instance segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep spectral methods are gaining traction for image segmentation, adapting traditional spectral techniques.
• Instance segmentation using deep spectral methods is underexplored, facing challenges with feature map channel noise.

Purpose of the Study:

• To enhance instance segmentation performance within deep spectral frameworks.
• To address the issue of noisy and uninformative channels in self-supervised feature maps.
• To propose a more suitable similarity metric for instance segmentation compared to the dot product.

Main Methods:

• Introduced Noise Channel Reduction (NCR) and Deviation-based Channel Reduction (DCR) modules for feature map channel pruning.
• Proposed a new similarity metric, Bray-curtis over Chebyshev (BoC), for affinity matrix construction.
• Evaluated methods on Youtube-VIS 2019 and OVIS datasets.

Main Results:

• NCR and DCR effectively reduced noisy and uninformative channels, improving segmentation accuracy.
• The BoC metric demonstrated superior performance over the dot product for instance segmentation.
• Significant improvements in mean Intersection over Union (mIoU) and instance segment quality were observed.

Conclusions:

• The proposed channel reduction techniques and BoC similarity metric substantially advance deep spectral instance segmentation.
• These methods offer a more robust and accurate approach to instance segmentation tasks.
• The findings pave the way for more effective deep spectral image analysis.