Robust Multi-generation Learned Compression of Point Cloud Attribute

Abstract: Existing learned point cloud attribute compression methods primarily focus on single-pass rate-distortion optimization, while overlooking the issue of cumulative distortion in multi-generation compression scenarios. This paper, for the first time, investigates the multi-generation issue in learned point cloud attribute compression. We identify two primary factors contributing to quality degradation in multi-generation compression: quantization-induced non-idempotency and transformation irreversibility. To address the former, we propose a Mapping Idempotency Constraint, that enables the network to learn the complete compression-decompression mapping, enhancing its robustness to repeated processes. To address the latter, we introduce a Transformation Reversibility Constraint, which preserves reversible information flow via a quantization-free training path. Further, we propose a Latent Variable Consistency Constraint which enhances the multi-generation compression robustness by incorporating a decompression-compression cross-generation path and a latent variable consistency loss term. Extensive experiments conducted on the Owlii and 8iVFB datasets verify that the proposed methods can effectively suppress multi-generation loss while maintaining single-pass rate-distortion performance comparable to baseline models.