Learning-Based Rate Control for Video-Based Point Cloud Compression.

Area of Science:

• Computer Vision
• Multimedia Signal Processing
• Data Compression

Background:

• Efficient rate control is crucial for Video-based Point Cloud Compression (V-PCC) due to transmission and storage limitations.
• Existing online updating methods for V-PCC rate control can be inaccurate due to inconsistencies in adjacent 2D frames.

Purpose of the Study:

• To propose a novel learning-based rate control method to enhance the rate-distortion (RD) performance of V-PCC.
• To improve the subjective quality and efficiency of dynamic point cloud compression.

Main Methods:

• Developed a low-latency synchronous rate control structure to minimize pre-coding overhead.
• Proposed a CNN-LSTM neural network for accurate prediction of basic unit (BU) parameters, replacing less reliable online updating.
• Introduced a patch-based clipping method for determining quantization parameters to prevent unnecessary clipping.

Main Results:

• The proposed CNN-LSTM network achieved accurate prediction of BU parameters.
• The patch-based clipping method effectively avoided unnecessary clipping during quantization parameter determination.
• Experimental results demonstrated superior RD performance compared to existing V-PCC rate control approaches.

Conclusions:

• The proposed learning-based rate control method significantly improves RD performance in V-PCC.
• The novel approach enhances subjective dynamic point cloud quality and overall compression efficiency.
• This method offers a more accurate and efficient solution for V-PCC rate control.