PaperlayerView exact revision on arXiv

Source-linked AI summary

Rate-Distortion Adaptive Primitive Selection for Omnidirectional Gaussian Splatting

Yulong Cheng, Youneng Bao, Junfeng Zhou, Mu Li, Jie Wen

arXiv:2609.34367v2cs.CV

TL;DR

Omnidirectional VR needs high-quality content with low decoding latency, but learned codecs can be slow and GS codecs do not generally select primitives by coding cost. OIC-GS combines hierarchical HEALPix primitives with rate-distortion optimization and a lightweight entropy model. It outperforms evaluated GS codecs, reducing WS-PSNR BD-rate by 49.6% over GaussianImage++ and 68.6% over SGI.

  • Problem

    Learned image codecs can make real-time VR decoding challenging, while GS codecs typically allocate primitives without jointly considering coding cost and distortion.

    Source: §1 · p. 1

  • Method

    OIC-GS uses hierarchical HEALPix-anchored primitives and a lightweight entropy model, jointly optimizing primitive existence under a spherical rate-distortion objective.

    Source: §1 · p. 2, §3.1 · pp. 3–4

  • Results

    49.6% lower WS-PSNR BD-rate than GaussianImage++ and 68.6% lower than SGI; OIC-GS outperforms tested GS codecs across four metrics and two benchmarks.

    Source: §4.2 · p. 7, §5 · p. 10

  • Takeaways & Limitations

    A single bitstream supports viewport-dependent, full-sphere, and progressive decoding, with the first viewport reaching final quality after 52% of the bitstream is decoded.

    Source: §1 · p. 2, §5 · p. 10

Abstract

from arXiv · show

Learned image codecs (LICs) achieve high reconstruction quality, but their decoding speed is often insufficient for immersive virtual reality (VR). Gaussian splatting (GS) codecs render much faster, yet still lag in reconstruction quality and typically decide primitive allocation without considering the coding cost of each primitive. We introduce OIC-GS, an omnidirectional GS codec with a new hierarchical HEALPix primitive grid representation. Gaussian primitives are anchored at predefined spherical locations, eliminating explicit coordinate coding. Finer levels refine their coarser ancestors, naturally supporting coarse-to-fine reconstruction and layered transmission. The predefined grid also enables efficient viewport decoding by selecting only view-relevant primitives. We further introduce a lightweight entropy model for quantized primitives and optimize the codec under a spherical rate-distortion objective. Primitives with insufficient rate-distortion benefit are automatically removed when their quantized opacity becomes zero, allowing OIC-GS to adapt both primitive density and level of detail without a fixed primitive budget. A single bitstream supports full-sphere, viewport-dependent, and progressive decoding. The first viewport reaches final quality after decoding only 52% of the bitstream, and is then rendered at 1,270 FPS. On a 100-image omnidirectional benchmark, OIC-GS outperforms all evaluated GS codecs, reducing WS-PSNR BD-rate by 49.6% over GaussianImage++ and 68.6% over SGI, which uses a learned entropy model.

1 INTRODUCTION

OIC-GS addresses inefficient omnidirectional GS coding and delayed viewport access with a hierarchical spherical representation optimized for rate-distortion. Its single bitstream supports viewport, full-sphere, and progressive decoding.

1 INTRODUCTION

  • LICs offer strong reconstruction quality but make real-time decoding challenging, while existing GS codecs tend to allocate primitives without jointly considering their coding costs.ERP's non-uniform sampling can cause excessive primitive allocation near the poles.Source: §1 · p. 1
  • Hierarchical HEALPix grids anchor primitives at predefined spherical locations, avoiding coordinate coding and ERP's sampling imbalance.Finer levels refine coarser ancestors, supporting coarse-to-fine reconstruction.Source: §1 · p. 2
  • The rate-distortion objective jointly accounts for primitive coding cost and distortion, enabling automatic selection as insufficiently useful primitives' quantized opacity falls to zero.A lightweight entropy model estimates primitive-attribute rates.Source: §1 · p. 2
  • One bitstream supports full-sphere, viewport-dependent, and progressive decoding; the first viewport needs only 52% of the file.The remaining content can be progressively delivered when needed.Source: §1 · p. 2
  • 49.6% lower WS-PSNR BD-rate than GaussianImage++; OIC-GS also outperforms tested GS codecs across bitrates.The hierarchical HEALPix representation supports coarse-to-fine coding.Source: §1 · p. 2

2 RELATED WORK

Omnidirectional codecs use planar projections, learned methods, or spherical representations, while Gaussian-splatting codecs have generally not used probability models to make primitive count rate-adaptive. OIC-GS renders viewports directly from spherical primitives.

2 RELATED WORK

  • Conventional omnidirectional compression uses projected panoramas with planar codecs or learned codecs, while other approaches use pseudocylindrical or spherical representations.Prior viewport-based methods decode a static image or a fixed chain of viewports before display, increasing latency at high resolutions.Source: §2 · p. 2
  • OIC-GS renders viewports directly from primitives at display resolution, avoiding a full-image network pass; equal-area HEALPix tiles do not expand polar views into full-width bands.Once decoded, the viewport is rendered in real time.Source: §2 · p. 3
  • Prior GS codecs generally optimize attributes for reconstruction quality rather than probability-model coding costs, so primitive counts are not driven by a rate gradient.GaussianImage++ densifies by reconstruction error, while LIG allocates by level of detail.Source: §2 · p. 3
  • SGI uses a learned entropy model, but transmits its location-conditioned hash grid and fixes primitive count in advance.OIC-GS instead conditions on the already reconstructed rendering state and lets code length drive primitive existence.Source: §2 · p. 3

3 METHOD

OIC-GS organizes panorama primitives on nested equal-area HEALPix grids and jointly fits their attributes and coding rates under a spherical rate-distortion objective.

3.1 HIERARCHICAL SPHERICAL GAUSSIAN REPRESENTATION

  • Each level places one primitive at each equal-area HEALPix cell center, with four finer cells per coarser cell and grid-fixed positions.The nested hierarchy supports coarse-to-fine coding.Source: §3.1 · p. 3
  • Quantized opacity controls primitive existence: intermediate states provide reduced contributions while neighbors and coarser levels adapt, and zero-opacity primitives fall back to coarser levels.Only active primitives are rendered with their own attributes and written to the bitstream.Source: §3.1 · p. 4, Figure 1 · p. 4
  • Coarse-to-fine rendering shares one feature space and ColorMLP, enabling full-sphere or viewport rendering and previews after partial levels.Each level refines the coarser state.Source: §3.1 · p. 5

3.2 ENTROPY AND RATE MODEL

  • RCC uses decoded coarse, parent, and even-neighbor contexts; ContextMLP predicts 3-component Gaussian mixtures for arithmetic coding, and the decoder reconstructs context identically.The feature rate charges active primitives, with buffered opacities assigned only part of their rate during training.Source: §3.2 · pp. 5–6, Figure 2 · p. 5

3.3 RATE-DISTORTION OPTIMIZATION

  • Joint fitting minimizes spherical-grid reconstruction error alongside rate and regularization, balancing each primitive’s distortion reduction against its code length.The active primitive count emerges from optimization rather than a preset budget.Source: §3.3 · p. 6

4 EXPERIMENTS

The experiments evaluate OIC-GS on a 100-image main benchmark and SUN360, finding consistent rate-distortion gains over tested GS codecs. They also assess viewport decoding and rendering, qualitative reconstruction, and the effect of design choices.

4.1 EXPERIMENTAL SETUPS

  • The evaluation uses 100 panoramas at 2048×1024 ERP resolution and reports WS-PSNR plus three viewport metrics on six 90°×90° views.The main benchmark is PNG; SUN360 supplies a separate 100-image cross-dataset evaluation, and BD-rate is reported against JPEG.Source: §4.1 · pp. 6–7, Table 1 · p. 7

4.2 RATE-DISTORTION PERFORMANCE

  • OIC-GS outperforms every tested GS codec at every bitrate on all four metrics across the main benchmark and SUN360.Against SGI, BD-rate falls by 68.6%, 67.1%, 57.7%, and 66.3% on WS-PSNR, V-PSNR, V-SSIM, and V-LPIPS; against GaussianImage++, it falls by 49.6% on WS-PSNR.Source: Figure 3 · p. 7, Figure 4 · p. 7, §4.2 · p. 7

4.3 DECODING RUNTIME AND QUALITATIVE EVALUATION

  • After decoding 52% of the file, the first viewport reaches final quality bit-exactly and renders at 1,270 FPS.Qualitative comparisons report the lowest bitrate among methods and better preservation of grass and foliage detail than GaussianImage and SGI.Source: §4.3 · p. 8, Figure 6 · p. 8

4.4 ABLATION STUDIES

  • Charging every active primitive its full bitrate increases equal-quality rate by 31.2%, the largest degradation among tested variants.A context-free prior increases rate by 5.9%, straight-through rounding by 2.5%, and other tested changes have negligible or at most 0.3% impact.Source: §4.4 · p. 9

5 CONCLUSION

OIC-GS uses hierarchical HEALPix grids to anchor primitives at predefined spherical locations, with an entropy model and quantized-opacity mask selecting primitives according to distortion benefit. The conclusion reports strong GS-codec rate-distortion results and viewport decoding from a partial bitstream.

5 CONCLUSION

  • OIC-GS anchors primitives on hierarchical HEALPix grids, avoiding coordinate coding while using an entropy model and quantized-opacity mask to retain primitives whose distortion reduction justifies their cost.Finer levels refine coarser ancestors.Source: §5 · pp. 9–10
  • The first viewport reaches final quality after decoding 52% of the bitstream, and OIC-GS reduces WS-PSNR BD-rate by 49.6% over GaussianImage++ and 68.6% over SGI.The single bitstream supports full-sphere, viewport-dependent, and progressive decoding.Source: §5 · p. 10

OVERVIEW OF THE APPENDIX

  • The appendix covers implementation and evaluation protocols, full operating points and cross-dataset comparisons, spherical bit allocation, decoding modes, ablations, resolution limits, and qualitative results.It also specifies the network architecture, training schedule, and rate and gradient conventions.Source: OVERVIEW OF THE APPENDIX · p. 14, A IMPLEMENTATION DETAILS AND… · p. 14

A.1 ARCHITECTURE AND TRAINING

The appendix specifies the model configuration and training conventions for OIC-GS, including its spherical hierarchy, quantization, level-specific coding schedule, and rendering-conditioned rate and gradient calculations.

A.1 ARCHITECTURE AND TRAINING

  • The hierarchy has seven primitive levels, ℓ∈{0,…,6}, with nside increasing from 4 to 256; a target grid uses nside 512 and 3,145,728 directions.Levels ℓ≤2 are always active; level 3 uses one coding pass, and levels {4, 5, 6} use a two-pass checkerboard schedule.Source: §A.1 · p. 14
  • Feature channels use 4-bit quantization with step 1/15, while opacity uses 2-bit quantization and a threshold τ=0.2 for the existence mask.The ContextMLP has about 3.2k parameters and is applied to each cell’s neighborhood context.Source: §A.1 · p. 14
  • The rendering and selection constants specify level-shared isotropic kernel widths, angular grid spacing, a learnable offset, and blending parameters.With ωlo=ωhi=τ, the four opacity states are charged 0, 1/3, 2/3, and 1 times their feature rate; ωlo controls the inactive-primitive STE slope.Source: §A.1 · p. 14, Table 4 · p. 15

A.2 RATE AND GRADIENT CONVENTIONS

  • Training hyperparameters are listed separately, while the context input concatenates its three parts with the level index.The appendix also explains coarse-level predictions and parent-cell indexing used by the context conventions.Source: §A.1 · p. 14, §A.2 · p. 15
  • The two-pass schedule codes even-indexed cells first and odd-indexed cells second, using decoded neighbors where available and coarse values otherwise.The same rule is applied in training, encoding, and decoding, supporting a round-trip check.Source: §A.2 · p. 15
  • The rate estimate charges each primitive according to its predicted feature code length, so the rate gradient pushes opacity down more strongly for primitives estimated to require more bits.For active primitives, distortion also contributes a gradient through their role in the weighted rendering average; the balance determines how primitives move among opacity states.Source: §A.2 · pp. 15–16

A.3 BASELINES AND METRICS

This appendix describes the baseline codecs and evaluation protocol, then reports operating-point variability and pairwise rate-distortion comparisons.

A.3 BASELINES AND METRICS

  • The comparison fits GS codecs per image using public implementations; JPEG2000 still outperforms the compared GS codecs, and parity with VVC intra is not claimed.OIC-GS is resampled to the 2048×1024 ERP grid with Lanczos-3, so it incurs resampling loss while ERP-fitted codecs are scored on their own grid.Source: §A.3 · p. 16

B ADDITIONAL RATE-DISTORTION AND CROSS-DATASET RESULTS

B.1 OPERATING POINTS AND PAIRWISE BD-RATES

  • At fixed λR, bitrate varies about 3× at the highest operating point and 4× at the lowest across images.The reported ranges are 0.56–1.76 bpp at the highest point and 0.07–0.26 bpp at the lowest.Source: §B.1 · p. 16
  • Pairwise BD-rates use each codec pair’s common quality range and are not directly comparable with the JPEG-anchored Table 1 values.For V-LPIPS, differing pairwise quality ranges reverse the SGI and GaussianImage order reported in Table 1.Source: §B.1 · p. 17, Table 7 · p. 17

B.2 CROSS-DATASET EVALUATION

The cross-dataset evaluation refits GS methods on SUN360, while the bit-allocation analysis compares where coded feature cost falls across spherical latitude bands.

B.2 CROSS-DATASET EVALUATION

  • SUN360 provides an independently collected 100-image test set, and the authors refit every GS method after verifying no overlap with the main dataset.Eight perceptual-hash candidate pairs were rejected by RGB comparison, with RMSE values of 34–67.Source: §B.2 · p. 17
  • On SUN360, OIC-GS beats every refitted GS codec at every tested bitrate on all four metrics, with larger margins than on the main benchmark.At 0.6 bpp on V-PSNR, its margins are +3.9 dB over GaussianImage, +4.1 over SGI, +6.9 over LIG, and +3.2 over GaussianImage++; JPEG anchoring may affect JPEG comparisons but not GS-to-GS comparisons.Source: §B.2 · p. 17
  • At 0.6 bpp on SUN360, the mean per-image V-PSNR margin over GaussianImage is +4.0 ± 0.21 dB across 90 images.The +4.0 dB is the mean of per-image margins; the mean rate-distortion-curve margin over those images is +3.9 dB.Source: §B.2 · pp. 17–18

C.1 MEASUREMENT PROTOCOL

  • The bit-allocation protocol assigns coded feature bits to latitude bands by primitive location and compares them with the corresponding coding costs of other codecs.The sphere is divided into 64 latitude bands; OIC-GS feature lengths use the arithmetic coder’s CDF table, while the mask, opacity streams, header, and MLP weights are excluded.Source: §C.1 · p. 18

C.2 BIT ALLOCATION OVER LATITUDE BANDS

  • The polar bands cover 14.2% of the sphere but receive 13.1% of OIC-GS feature bits, versus 27.9–34.4% for ERP-based codecs.ERP-based bit shares track the 34.4% ERP pixel share; the remaining OIC-GS feature-bit density variation reflects image content.Source: §C.2 · p. 18

D BITSTREAM FORMAT AND DECODING MODES

The appendix evaluates full-sphere, viewport-dependent, and progressive decoding from OIC-GS bitstreams. It reports bit-exact viewport recovery, fast rendering, and mechanisms for skipping or progressively receiving data.

D.1 BITSTREAM FORMAT AND CODEC VERIFICATION

  • The bitstream stores level-wise feature streams, masks, and opacity symbols; round-trip tests recover quantized features and opacities bit-exactly.Feature streams account for 38% and 68% of file size at the lowest and highest operating points, respectively; viewport-dependent decoding matches full decoding pixel-for-pixel.Source: §D.1 · p. 19

D.2 MEASUREMENT PROTOCOL AND RUNTIME DEFINITIONS

  • Fixed-viewport rendering reaches 1,270 FPS, while full-sphere rendering runs at 140 FPS on the reported GPU.The fixed-viewport frame takes 0.79 ms; the ColorMLP accounts for 60% of GPU time, and host copying is excluded from the benchmark frame rate.Source: §D.2 · pp. 19–20

D.3 TILE-ALIGNED RANDOM ACCESS

  • Tile alignment makes viewport data independently skippable, allowing 39–50% of a file to be skipped with 1.2% average storage overhead.The 4,096-cell NESTED tiles store mask, opacity, and feature segments as contiguous byte ranges; equal-area tiles avoid expanding a polar viewport into a full-width band.Source: §D.3 · pp. 20–21, Table 9 · p. 21

D.4 FULL-SPHERE AND VIEWPORT-DEPENDENT DECODING

  • Viewport decoding selects streams intersecting the viewport and its dependencies, and is bit-exact with full decoding inside the viewport.The padded 90°×60° frustum covers about 13% of target directions, although shared coarse levels and context dependencies reduce the fraction of file that can be skipped.Source: §D.4 · p. 21

D.5 PROGRESSIVE STREAMING AND HEAD ROTATION

  • Progressive streaming displays the first viewport after receiving 52% of the file, while level-aligned prefixes provide a full-sphere preview from the first 10%.After the remaining tiles arrive, the sphere completes without changing any first-viewport pixel; the two head-rotation strategies average 314/891 FPS across the reported sessions.Source: §D.5 · p. 21, §D.6 · p. 22, Table 11 · p. 22

D.6 DEPLOYMENT CONSIDERATIONS

  • Both head-rotation strategies stay within the 11.1 ms frame budget of a 90 Hz display, and static tile offsets support viewport-adaptive HTTP range requests.The tile-aligned format adds 1.2% average storage overhead and requires neither server-side transcoding nor per-client bitstreams.Source: §D.6 · p. 22

E.1 ESTIMATED BITRATE VERSUS WRITTEN FILE

  • Written feature bytes exceed the masked feature-rate estimate by 1.9% on average and 0.7% at the median.The difference is attributed to termination overhead from the 32 parallel arithmetic-coding streams; the weighted rate estimate is used in training, not as the reported written bitrate.Source: §E.1 · p. 22

E.2 ABLATION VARIANTS AND DESIGN CONSTANTS

The ablations compare rate-aware opacity weighting, context modeling, and rate estimation choices; the resolution appendix distinguishes the default codec configuration from an additional finer level.

E.2 ABLATION VARIANTS AND DESIGN CONSTANTS

  • On SUN360, removing rate-aware opacity weighting increases relative rate by 29.0%/28.1% for WS-PSNR/V-PSNR; removing RCC increases it by 6.5%/7.9%.Using STE instead of uniform noise yields +2.4%/+3.3% for WS-PSNR/V-PSNR; the design-constant variants range from −0.6% to +0.5%.Source: §E.2 · p. 23
  • Without rate-aware opacity weighting, the model is worse on 96 of 100 main-benchmark images and retains 6.1× as many primitives, spending 0.66 instead of 0.29 bpp.This variant remains calibrated: written feature bytes differ from the estimate by 2.5%, similar to the full model.Source: §E.2 · p. 23
  • Without RCC, the variant is worse on 98 of 100 main-benchmark images and selects 0.84× as many primitives as the full model.The text attributes the lower selection to a context-free prior raising predicted average primitive cost, which pushes opacities down more strongly at fixed λR.Source: §E.2 · p. 23
  • Replacing uniform noise with STE leaves quality within 0.02 dB WS-PSNR on the main benchmark, while the loss appears in bitrate.This variant is worse than the full model on 96 of 100 main-benchmark images.Source: §E.2 · pp. 23–24

F RESOLUTION LIMIT AND THE L7 EXTENSION

  • The L7 configuration adds level 7 at nside=512 to the default L6 model, which ends at level 6 and nside=256.The paired resolution analysis reports native and end-to-end score improvements on all 10 scenes, but its fits use continuous features and all primitives active, not matched-bitrate codec operating points.Source: F RESOLUTION LIMIT AND THE L7 EXTENSION · p. 24, Table 12 · p. 24

F.1 CAPACITY ANALYSIS WITHOUT RATE

Adding level 7 improves reconstruction under a fixed primitive capacity, with gains on every tested scene. Its end-to-end gains are smaller than native-sphere gains because ERP evaluation includes resampling loss.

F.1 CAPACITY ANALYSIS WITHOUT RATE

  • The paired L6/L7 comparison fits 10 scenes for 10k iterations with all primitives active and λR=0, making it a fixed-capacity analysis rather than a bitrate-matched comparison.The fits contain 1,048,512 L6 primitives and 4,194,240 L7 primitives.Source: §F.1 · p. 24
  • +8.502 dB native sphere PSNR and +2.752 dB end-to-end WS-PSNR on average across all 10 scenes show level 6 is a capacity limit under this setting.The reference ERP-to-HEALPix-to-ERP round trip reaches only 34.882 dB; the fits' gap to it shrinks from 3.808 to 1.056 dB.Source: §F.1 · p. 24
  • At nside=512, level 7 improves end-to-end ERP WS-PSNR for every selected image by +2.752 dB on average, with gains ranging from +2.035 to +3.682 dB.Native L7 gain correlates with lower reference round-trip fidelity (Pearson correlation −0.739) and global and nadir high-frequency energy (+0.642 and +0.721); average nadir gain is +10.480 dB versus +8.022 dB at the equator.Source: Figure 9 · p. 25
  • ERP-based scores include two resamplings that reduce the apparent gain; the additional level's bitrate value is evaluated separately with rate, quantization, and selection enabled.On the main benchmark, operating points score 1.9–2.8 dB higher in the native domain than end-to-end.Source: §F.1 · p. 25

F.2 RATE-DISTORTION EVALUATION WITH L7

With rate, quantization, and selection enabled, adding level 7 improves rate-distortion performance on the evaluated subset. Only a small fraction of its cells remain active, while selection also redistributes primitives across levels.

F.2 RATE-DISTORTION EVALUATION WITH L7

  • L7 is evaluated on a 20-image subset with the same selection mechanism and target grid as L6, but takes 18.0 versus 12.7 minutes per image and has greater run-to-run variation.The authors retain L6 as the default and report L7 as a scalability result.Source: §F.2 · p. 25
  • 7.7% lower WS-PSNR BD-rate and 8.2% lower V-PSNR BD-rate for L7 relative to L6, with a +0.24 dB gain at matched bitrate on all 20 images with defined BD-rate.The comparison uses the common quality range for each pair at three operating points.Source: §F.2 · p. 25
  • The median image selects only 0.14–0.56% of level-7 cells across the three bitrates, with a per-image range of 0.02–1.87%.The median level-7 selection ratio increases 4.0× as λR decreases by 2.2×.Source: §F.2 · p. 26, Table 14 · p. 26
  • RDPS reduces active level-6 primitives by 28–56%, while total active primitives stay unchanged at the lowest bitrate and fall 9–15% at the two highest.Levels 4–5 retain slightly more primitives, so selection redistributes them across levels rather than simply adding more.Source: §F.2 · p. 26
  • Selected level-7 cells concentrate on high-frequency content: 65–88% fall in the top gradient quartile, versus 1.9–3.2× the image-mean gradient for level 6.The median selection ratio rises as λR decreases, and 45% of level 7 remains active after the 3k-iteration warm-up before the rate term selects among them.Source: §F.2 · p. 26
  • On scene 87, active level-7 cells cluster on foliage, rocky ground, and shoreline, not sky or open water; lowering λR selects more cells without changing their locations.Source: Figure 10 · p. 27

G ADDITIONAL QUALITATIVE RESULTS

Across five selected scenes and six viewing directions, OIC-GS generally performs strongly against GS codecs, COIN, and JPEG. The qualitative comparisons also show that its advantage is most visible in textured regions.

G ADDITIONAL QUALITATIVE RESULTS

  • A window where OIC-GS outperforms JPEG appears in 28 of 30 panels, and every baseline uses at least as many bits as OIC-GS on every displayed scene.The five scenes span the 10th through 90th percentiles of OIC-GS's per-image WS-PSNR gain over SGI (+1.2–+3.0 dB) and were selected before inspecting results.Source: G ADDITIONAL QUALITATIVE RESULTS · p. 28
  • The qualitative comparison shows little difference in largely uniform sky regions, while textured grass and buildings show more recovered high-frequency detail with OIC-GS.Source: G ADDITIONAL QUALITATIVE RESULTS · p. 28
  • OIC-GS has the highest viewport PSNR among GS codecs and COIN in 25 of 30 panels and the highest patch PSNR in 27.It matches or exceeds JPEG2000 in 11 of 12 equatorial panels across scenes 7, 32, and 55, while JPEG2000 leads at every scene's nadir.Source: Figure 11 · p. 28
  • On scenes 74 and 46, every baseline uses more bits than OIC-GS; four GS-codec viewport-PSNR exceptions are GaussianImage++ at 30–82% higher bitrate and by 0.1–0.6 dB.The exceptions are the yaw-180° and nadir views of scene 74 and the zenith and nadir views of scene 46.Source: Figure 12 · p. 29
  • On scenes 7 and 32, OIC-GS has the highest viewport PSNR among GS codecs and COIN in all 12 panels, while every baseline uses more bits.Their OIC-GS bitrates are 0.527 and 0.581 bpp, respectively.Source: Figure 13 · p. 30
Loading 2609.34367v2…