RSTSIC: Reparameterized Swin Transformer Stereo Image Compression

Official implementation of RSTSIC: Reparameterized Swin Transformer Stereo Image Compression

Key Features

• State-of-the-Art Performance: Outperforms traditional codecs and learning-based compression methods on PSNR and MS-SSIM metrics.
• High Efficiency: Structural reparameterization reduces inference complexity without sacrificing performance.
• Robust Generalization: Validated on both urban (Cityscapes) and indoor (InStereo2K) stereo datasets.
• Real-Time Ready: Low latency and lightweight design.

Evaluation Results

Qualitative Results

Ground Truth

RSTSIC bpp=0.043, PSNR=33.8

LDMIC bpp=0.045, PSNR=33.4

SASIC bpp=0.052, PSNR=32.8

Environment Configuration

Our code was tested with the following environment configurations. It may work with other versions.

• Ubuntu 20.04
• NVIDIA Tesla T4 GPU
• CUDA 12.4
• Python 3.9
• PyTorch 2.1.0 + cu121
• CompressAI 1.2.0

Ackownledgement

Our code is based on the implementation of CompressAI. We thank the authors for open-sourcing their code.