We present a GPU-native volumetric-video system for real-time reconstruction of graphics-compatible textured triangle meshes from multiple RGB-D sensors. Existing TSDF systems accelerate fusion and mesh extraction, but extending them to per-frame textured meshing adds CPU mesh processing, CPU-side texturing work, or extra CPU-GPU transfers between stages. After input staging, our method keeps the reconstruction and texturing stages (TSDF fusion, Marching Cubes, connected-component noise removal, view-based texture mapping, inpainting, and radiometric seam correction) resident on the GPU; no intermediate mesh or image data returns to the host between these stages. A fixed-size-array representation of mesh topology gives constant-time neighbor access on the GPU and makes the downstream mesh-level operations data-parallel. Because we texture each face from a single camera, exposure seams that vary smoothly across each image remain; we estimate a per-camera low-frequency gain field and bake it into the atlas on the GPU. On the CMU Panoptic dataset, our method is faster than the baselines at every tested voxel resolution and number of sensors, achieves peak speedups of more than an order of magnitude over CUDA-accelerated nvblox and CPU-based Open3D, and approaches an offline textured-meshing baseline in perceptual re-rendering quality.
Figures & tables
Figure 1: Per-frame processing time versus voxel size on the office2 sequence of CMU Panoptic [ 21 , 22 ] with five Kinect inputs. Both axes use log scale. “Ours” includes input upload, preprocessing, and radiometric seam correction ( Sec. 3.6 ) and remains real-time at 0.02m . Open3D (CPU) is run without noise removal here and in Fig. 4 ; Tab. 1 includes it.
Figure 2: Our GPU-native textured-meshing pipeline. Each frame uploads the raw depth and color images in one batch; Preprocessing undistorts them and registers depth to color on the GPU. Apart from a few scalar counters and flags only the final mesh, UV coordinates, and corrected atlas return to the CPU. # n gives the main CUDA kernel launches per stage ( K cameras); Nnr and Nip are the label-propagation and inpainting iterations (capped at 100 and 5 ) and Nsc=55 . Red stars mark the stages that build on the fixed-size-array mesh topology ( Sec. 3.2 ); the seam correction ( Sec. 3.6 ) reuses this adjacency and writes the corrected atlas before output.
Figure 3: Radiometric seam correction. Left: cameras a and b see the same surface at different exposures (dark in image a , light in image b ); faces above the seam are textured from a and faces below it from b , so brightness mismatches along the texture seam . Center: both images are sampled at eight corresponding positions per seam edge (four drawn), each inset toward the opposite vertex of that camera’s own face ( ∘ edge point, ∙ sample); the median log-luminance difference ds constrains the per-camera log-gain grids gk , which live in each camera’s image space. The cells pas,pbs at the seam midpoint p are coupled by ga[pas]−gb[pbs]≈ds , smoothness spreads the correction to neighboring cells, and eg^k multiplies the camera image (red brightens, blue darkens, as in Fig. 7 ). Right: the corrected images and mesh; the correction is strongest at the seam and fades with distance, and the mismatch at p is reduced. Two cameras are shown; with K cameras each seam couples the grids of its two cameras in the same way, and all K grids are solved jointly.
Figure 4: Per-frame processing time versus the number of sensors. Measured on the office2 sequence of CMU Panoptic at a voxel size of 0.02m . The vertical axis is the processing time per frame (ms, log scale). Our method’s processing time increases more slowly with the number of sensors than the baselines. “Ours” is the full method, including the radiometric seam correction stage ( Sec. 3.6 ); it stays below all baselines at every sensor count.
Stage
Ours
Ours w/o sc
Ours (plain)
nvblox
Open3D (CPU)
Open3D (CUDA)
CPU → GPU
2.0
1.9
1.9
3.3
N.A.
4.4
Preprocessing †
0.3
0.3
0.3
N.A.
N.A.
N.A.
TSDF Fusion
0.4
0.4
0.4
36.3
95.7
8.2
Marching Cubes
0.5
0.5
0.6
28.1
299.6
5.7
Noise Removal
3.2
2.9
3.4
N.A.
957.0
N.A.
Texture Mapping
0.2
0.1
–
N.A.
N.A.
N.A.
Table 1: Per-stage runtime breakdown (ms) on office2 (voxel 0.02m , five Kinect inputs; about 0.59 M output faces per frame). “Ours w/o sc” disables the seam correction, and “Ours (plain)” drops all texture stages for comparison with the per-vertex-color baselines nvblox, Open3D (CPU; Scalable variant with noise removal), and Open3D (CUDA). Inpainting includes the adjacency rebuild (1.3 of 1.7 ms). Columns are separate runs; shared stages vary by up to ∼0.3 ms with GPU timing.
Figure 5: Qualitative re-rendering comparison on the office2 and band2 sequences of CMU Panoptic (camera 9, frames 933 and 403, respectively). Meshes reconstructed from five Kinect inputs at a voxel size of 0.02m are re-rendered from the RGB viewpoints of a different set of five Kinects that were not used as input. “Ours” includes the radiometric seam correction ( Sec. 3.6 ). Compared with the per-vertex-color outputs of Open3D and nvblox, our method reproduces facial and clothing textures more clearly, approaching the offline result at real-time speed. Quantitative scores are in Tab. 2 .
office2
band2
Method
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Open3D w/o nr
16.483
0.783
0.353
17.635
0.807
0.329
Open3D w/ nr
16.408
0.784
0.353
17.590
0.808
0.330
nvblox
15.604
0.759
0.378
16.549
0.785
0.355
Ours (vtx color)
16.160
0.787
0.337
17.320
0.808
0.319
Ours w/o sc
15.869
0.771
0.299
17.115
0.796
0.277
Table 2: Quantitative re-rendering quality on office2 and band2 (held-out five-Kinect evaluation, voxel 0.02m ), measured against the GT images with 5% trimmed mean PSNR, SSIM [ 47 ] , and LPIPS [ 50 ] . “Ours” is the full method including radiometric seam correction and “Ours w/o sc” the ablation; “Ours (vtx color)” replaces the texture atlas with per-vertex color on the same geometry, isolating the atlas contribution. Metrics are computed on the common valid-pixel mask, the intersection of covered pixels across all eight evaluated variants. The qualitative comparison is in Fig. 5 .
Figure 6: Effect of noise removal and inpainting on a foreground-only in-house capture. The subject swings a shirt, producing noisy depth near the cloth and hands. (a) uses texture mapping without noise removal or inpainting; it leaves isolated fragments and untextured faces. (b) Noise removal suppresses the fragments, and (c) inpainting fills many remaining untextured faces. See the supplementary video for temporal behavior.
Figure 7: Radiometric seam correction on office2 , rendered from a held-out viewpoint. (a) Full view without and with correction, and the source-camera mosaic, where each color marks the input camera that textures that face, so inter-camera seams lie at the color boundaries. (b) Two zoomed regions, a wall-panel junction (cyan) and the floor (red), each as GT, without, and with correction. (c) Each input camera’s 24×16 gain field exp(gk) (red brightens, blue darkens). Seam residuals are in Tab. 3 .
office2
band2
Seam residual ↓
Constr.
Output
Constr.
Output
Before (Ours w/o sc)
0.159
0.162
0.151
0.150
After (Ours)
0.101
0.108
0.103
0.109
Reduction
36.5%
33.3%
31.8%
27.3%
Table 3: Mean absolute log-luminance mismatch at the seams between input cameras ( 0.02m , frames of Tab. 2 ), inside the solver (Constr.) and on the written atlas (Output).
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Texture atlas of our output ( office2 , frame 445, 0.02m , seam correction on). (a) The atlas is the stack of the five input color images, one band per Kinect, with the per-camera gain baked in. (b) The mesh textured from it, rendered from a held-out camera. (c) A crop of the top band with the UV triangles of the faces that sample it: UVs are pixel coordinates in the source image, so the mesh reads the texture exactly where the camera saw it.
Stage / parameter
Value
Reconstruction
Voxel size
0.02m
SDF truncation band
4× voxel size
Mesh post-processing
Noise removal: min. component
500 faces
Noise removal: max. iterations
100
Appendix
Table 4: Default hyperparameters for the Panoptic evaluation of Sec. 4 . Overrides for ablations and the auxiliary capture are stated in their respective sections.
Figure 9: Seam-residual sampling geometry ( Sec. 3.6 ). A seam is the edge vivj shared by face a (textured from camera a ) and face b (camera b ). For each of the N=8 sample points e at t=(i+21)/N along the edge, both faces are sampled at corresponding shared-edge positions e , and the image-space sample position is inset by δ from e toward the projection of that face’s opposite vertex ( oa or ob ), not along the edge normal. The inset is applied independently in each camera’s projected coordinates, with δ=0.5 proxy pixels ( =4 full-resolution pixels at the 1/8 proxy), moving the samples toward the interiors of their respective projected faces. The inset positions need not correspond to the same 3D point. Dark or invalid samples are rejected, and the seam residual ds is the median over the remaining samples of logYb−logYa ; seams whose faces fold sharply ( na⋅nb<0.5 ) or project degenerately are discarded.
office2
band2
Variant
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
off
15.869
0.7714
0.2988
17.115
0.7960
0.2773
constant ( 1×1 )
15.946
0.7786
0.2932
17.136
0.7984
0.2737
luminance
15.976
0.7847
0.2876
17.174
0.8038
0.2701
RGB
15.977
0.7847
0.2873
17.176
0.8039
0.2696
Appendix
Table 5: Luminance-only vs. per-channel RGB seam correction (held-out, 0.02m , 100 frames, common mask over all eight evaluated variants). “off” is no seam correction; “constant” solves one luminance gain per camera ( 1×1 field) and is scored on the same masks.
Variant
PSNR ↑
SSIM ↑
LPIPS ↓
off
22.645
0.9404
0.1024
luminance
22.842
0.9509
0.0948
RGB
22.847
0.9513
0.0941
clean input
22.988
0.9516
0.0937
Appendix
Table 6: Synthetic capture (held-out cameras, 0.01m , 150 frames, shared mask). “clean input”: reconstruction from the unperturbed renders, the upper bound.
Figure 10: Synthetic capture. (a) Without correction the floor and the shirt split into brighter and darker patches where the texture switches between cameras; both fields remove them. (b) The correction is relative: it reproduces the differences between cameras across the seams, not each camera’s applied gain, which is why the two rows differ in amplitude.
Figure 11: Luminance vs. RGB seam correction on the in-house five-camera capture (turntable view from behind; both with ∣g∣≤0.4 and λsm=5 ). (a) Without correction, with the luminance-only field, with the per-channel RGB field, and the source-camera mosaic: the red and yellow cameras texture most of the back, and their boundary crosses the T-shirt (cyan box) and the plaid shirt at the waist (red box). (b) Zoom-ins of the boxed regions: the darker, greenish T-shirt patch of the yellow camera and the brightness step on the plaid shirt are equalized by both variants. (c) Input views (background masked) and the solved gain fields: both brighten the yellow camera by up to about 1.5× toward the center of its view, and the three channels of the RGB field nearly coincide. The fields stay smooth and low-frequency and shift each camera’s colors as a whole.
office2
band2
Texturing
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Offline (mvs)
13.303
0.674
0.390
14.605
0.704
0.360
Ours
15.290
0.761
0.311
16.535
0.782
0.289
Appendix
Table 7: Offline texturing (mvs-texturing) applied to our geometry vs. our real-time texturing, on identical meshes (held-out, 0.02m , frames 400 – 429 ).
Voxel
PSNR ↑
SSIM ↑
LPIPS ↓
Δ PSNR
Δ LPIPS
ms/frame
GPU mem
0.02m
15.821
0.795
0.271
–
–
13.1
2.7 GB
0.01m
15.704
0.768
0.308
−0.12
+0.038
34.6
5.7 GB
0.005m
14.571
0.625
0.445
−1.25
+0.175
138.9
24.1 GB
Appendix
Table 8: Voxel size versus held-out re-rendering quality, runtime, and GPU memory on office2 (same pipeline, shared mask, frames 400–547 at stride 3, cameras 1,3,5,7,9, 10cm depth tolerance). Δ columns are relative to the 0.02m row, from unrounded values. Runtime is the per-frame processing time (low-95% mean) over 2000 frames (400–2399); the 0.02m row uses the same five-input timing trace as Tab. 1 . Memory is the peak device memory over these 2000 frames (static grid and mesh buffers plus per-frame allocations). Quality uses the 50-frame subset specified above.
Figure 12: Voxel size versus reconstruction detail on office2 , held-out camera 7, frame 445 (crop). Finer voxels sharpen silhouettes (chair, table rim) but leave more small holes and noise on the wall and floor, where each voxel receives fewer depth samples; the truncation band also shrinks with the voxel (four voxels: 8, 4, and 2cm ), so voxel size is not the only change. Overall the held-out metrics decrease (see text). Thin, dark structures such as the stool remain missing at every setting because the depth sensors do not observe them reliably.
τocc
PSNR ↑
SSIM ↑
LPIPS ↓
seam res. (before → after)
office2
1.1m (loose)
14.882
0.737
0.327
0.328 → 0.271
10cm (default)
15.556
0.765
0.304
0.158 → 0.104
band2
1.1m (loose)
16.110
0.753
0.310
0.287 → 0.240
10cm (default)
16.521
0.781
0.288
0.152 → 0.104
Appendix
Table 9: Occlusion threshold on office2 and band2 (same build, shared mask per scene, frames 400–547 at stride 3, cameras 1,3,5,7,9). Rejecting occluded projections removes the ghost texture and improves all three held-out metrics; the seam residual columns are the seam constraint residual of Tab. 3 before and after seam correction.
Figure 13: Occlusion-aware view selection on office2 (held-out camera 7, frame 405, crop). With the loose 1.1m tolerance, the green chair’s texture spills onto the floor and the table’s white texture onto the chair. The 10cm test rejects these occluded projections; inpainting fills the affected faces. Both use the same reconstruction settings.
Figure 14: Application with SfM+MVS-style input. (a) Point-cloud output of VGGT [ 46 ] . (b) Textured mesh obtained by applying our method to the VGGT-derived geometry and camera views.