Heatmap Regression via Randomized Rounding.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Heatmap regression is standard for semantic landmark localization in deep learning.
• It struggles with sub-pixel localization due to quantization error, limiting accuracy.
• Existing solutions often increase computational cost via high-resolution heatmaps.

Purpose of the Study:

• To address the sub-pixel localization problem in heatmap regression.
• To develop a method that improves localization accuracy without increasing computational cost.
• To propose an unbiased and lossless quantization system for heatmap regression.

Main Methods:

• Formal analysis of quantization error in vanilla heatmap regression.
• Introduction of a novel quantization system using randomized rounding.
• Probabilistic encoding of fractional coordinates during training and decoding during testing.

Main Results:

• The proposed quantization system is proven to be unbiased and lossless.
• Demonstrated effectiveness on facial landmark localization (WFLW, 300W, COFW, AFLW) and human pose estimation (MPII, COCO) datasets.
• Achieved efficient and accurate semantic landmark localization.

Conclusions:

• The novel quantization system effectively overcomes sub-pixel localization limitations in heatmap regression.
• The method offers a balance between localization accuracy and computational efficiency.
• This approach advances deep learning-based landmark localization techniques.