Louis Fréneau

Video-based Point Cloud Compression (V-PCC) efficiently compresses dynamic point clouds by projecting 3D patches into 2D occupancy, geometry, and attribute maps before video encoding. The reference attribute map generation pipeline relies on assumptions inherited from the standardization process, notably that global image smoothing and tighter patch packing improve coding performance. This work systematically reassesses these assumptions under practical encoding constraints using uvgVPCCenc, an open source V-PCC framework designed for real-time encoding. Rather than optimizing individual map generation stages independently, we adopt a holistic approach that analyzes the interactions between attribute background filling, attribute map conversion, and patch packing. To this end, we examine three straightforward methods: a local block-based background filling algorithm, a local RGB444-to-YUV420 attribute map conversion, and patch spacing. Their impact is evaluated on both uvgVPCCenc and the V-PCC reference encoder TMC2 using coding efficiency, computational complexity, and subjective visual quality. The results show that these low-complexity methods can substantially reduce map generation complexity. The proposed background filling accelerates the overall encoding by up to 1.41 \(\times\) , while the proposed attribute map conversion increases this speedup to 1.50 \(\times\) and eliminates visual artifacts caused by the interaction between background filling and the reference conversion. Furthermore, moderate patch spacing improves reconstructed quality by mitigating inter-patch color bleeding with only a marginal bitrate increase. Overall, our findings demonstrate that several assumptions underlying the reference V-PCC map generation pipeline no longer hold under practical encoding constraints. More importantly, it demonstrates that evaluating map generation stages in isolation can lead to misleading conclusions, advocating a holistic methodology for the design of future practical V-PCC encoders.