3DTexMOR: 3D Gaussian Multi-Object Removal via Texture-Space Inpainting

• 3D object removal reconstructs a scene, removes specified objects, and completes their occluded regions with plausible geometry and appearance.Source: §1 · p. 1
• Complex multi-object scenes limit the context visible in individual views, while inconsistent inpainted appearances can create conflicting supervision and blurry 3D reconstructions.Source: §1 · p. 2
• 3DTexMOR improves PSNR by over 5.8 dB and reduces LPIPS by at least 22% versus the state-of-the-art baseline on synthetic diffuse multi-object scenes.Source: §1 · p. 2
• 3DTexMOR combines complementary multi-view observations in unified texture space, then inpaints and reprojects completed maps to provide shared supervision for scene completion.Source: §1 · p. 2
• The framework aggregates intrinsic images into texture maps and uses geometrically regularized Gaussian completion to constrain recovered 3D structure.Source: §1 · p. 2

• LERF and SA3D (Cen et al., 2023) extend language features or SAM prompts to NeRF, while Feature 3DGS and LangSplat distill semantic features into Gaussians.Source: §2 · p. 2
• Gaussian Grouping attaches segmentation or object identity features to Gaussians, enabling object grouping, selection, and local editing.Source: §2 · p. 3
• NeRF- and 3DGS-based removal methods extend 2D inpainting to scenes, including diffusion-guided Gaussian editing and intrinsic-space inpainting.Source: §2 · p. 3
• Most existing methods still rely on image-based inpainting and mainly target single-object removal, making complex multi-object scenes difficult to complete consistently in geometry and appearance.Source: §2 · p. 3

3.1 OVERVIEW OF THE 3DTEXMOR FRAMEWORK

• The framework removes selected objects and completes occluded regions by combining shared intrinsic-texture supervision with geometric constraints on completion Gaussians.It retains the original Gaussians for unremoved content and optimizes completion Gaussians for the missing regions.Source: §3.1 · p. 3

3.2 OBJECT-IDENTITY AND INTRINSIC 3DGS

• Object identity features enable selective removal, while intrinsic attributes provide view-independent appearance for aggregation across views.Appearance is decomposed into view-independent diffuse and view-dependent glossy components rather than aggregating RGB observations directly.Source: §3.2 · pp. 3–4

3.3 TEXTURE-SPACE INPAINTING

• Multi-view intrinsic observations are mapped to a shared texture space, inpainted, and reprojected to provide consistent supervision across camera views.The space assumes completion regions lie approximately on an estimated support plane; bilinear and viewing-angle weights combine observations, and diffuse, Fresnel, and roughness maps are inpainted independently.Source: §3.3 · pp. 4–5

3.4 GEOMETRICALLY REGULARIZED GAUSSIAN COMPLETION

• Completed textures initialize Gaussians on the estimated support plane, whose geometry is softly regularized during optimization.The regularizers penalize departure from the plane, drift from initial texture coordinates, and changes from initial scales.Source: §3.4 · pp. 5–6

3.5 LOSS FUNCTIONS

• Completion Gaussians are optimized with intrinsic supervision and geometric regularization, including penalties for plane deviation, coordinate displacement, and scale change.Rendered intrinsic maps are compared with reprojected targets within the camera-view inpainting mask.Source: §3.5 · p. 6

4.1 IMPLEMENTATION DETAILS

• On an RTX 4090D, initial 3DGS reconstruction takes about 40 minutes, texture-space inpainting about 30 seconds, and Gaussian completion about 4 minutes.On Syn-Reflective, full physically-based rendering reaches 40.0 FPS for approximately 30,000 Gaussians at 800 × 800 resolution.Source: §4.1 · p. 6

4.2 EXPERIMENTAL SETUP

• Evaluation covers synthetic reflective and diffuse scenes and real-world single- and multi-object scenes, using full-image and mask-domain PSNR, SSIM, LPIPS, and FID metrics.The synthetic datasets each contain three scenes; Real-Single contains three scenes and Real-Multiple four.Source: §4.2 · pp. 6–7, Table 1 · p. 7

4.3 COMPARISONS WITH PRIOR WORKS

• 3DTexMOR achieves the best performance across reported metrics on Syn-Diffuse; on Syn-Reflective, it beats three baselines on PSNR, SSIM, and FID, while GOR-IS has slightly lower LPIPS.Visual comparisons also show fewer artifacts and less blurriness on real-world and synthetic scenes.Source: Table 1 · p. 7, Figure 4 · p. 7, §4.3 · p. 8

4.4 ABLATION STUDIES

• Replacing texture-space inpainting with image-space inpainting produces more artifacts in occluded regions that propagate into renderings; texture-space inpainting improves PSNR by approximately 7 dB across the synthetic datasets.The comparison averages metrics over Syn-Diffuse and Syn-Reflective.Source: §4.4 · p. 8, Table 2 · p. 8, Figure 5 · p. 8
• On reflective scenes, intrinsic decomposition removes deleted-object reflections and improves all reported metrics, while its improvement is less pronounced on diffuse scenes.The decomposition provides view-independent attributes for reducing glossy-surface artifacts.Source: §4.4 · pp. 8–9, Figure 6 · p. 9
• The anchor loss constrains drift along the estimated plane, while the surface loss penalizes perpendicular distance; adding both progressively improves performance, with the full model achieving the best metrics.The losses produce more uniform normals and visibly improve completed-region renderings.Source: §4.4 · p. 9, Table 3 · p. 9

5 CONCLUSION

• 3DTexMOR outperforms recent Gaussian object removal methods on synthetic and real-world scenes.The framework aggregates complementary multi-view observations in shared texture space and adds geometric constraints to recover plausible structure.Source: §5 · p. 9
• A common texture space combines complementary multi-view observations to provide richer information and reduce appearance inconsistencies during completion.Geometrically regularized Gaussian completion complements texture supervision with constraints on the completed regions.Source: §5 · p. 9
• Completion regions must be well approximated by a common plane; planar mapping may misalign reprojected supervision on curved, deformable, or irregular surfaces.Extending texture-space representation to more general surfaces is identified as future work.Source: §5 · p. 9

A TECHNICAL DETAILS

• The appendix details intrinsic attributes and the reflection renderer used in the representation.Source: A TECHNICAL DETAILS · p. 13
• It describes texture-space construction, inpainting, and reprojection.Source: A TECHNICAL DETAILS · p. 13
• It specifies Gaussian initialization and explains how intrinsic supervision and geometric regularization guide completion.Source: A TECHNICAL DETAILS · p. 13

A.1 INTRINSIC DECOMPOSITION AND REFLECTION RENDERING

• Each Gaussian carries diffuse radiance, normal-incidence Fresnel reflectance, and roughness; deferred rasterization produces their per-view maps and a normal map.Diffuse radiance includes diffuse illumination under fixed capture lighting and is not an estimate of unlit albedo.Source: §A.1 · p. 13
• The renderer traces reflected rays through the Gaussian scene for near-field radiance and uses an optimizable environment map for visible far-field radiance.Removing an object can change reflected radiance at a different surface location because the near-field term queries the scene.Source: §A.1 · p. 13
• Schlick’s approximation converts intrinsic Fresnel reflectance into an angle-dependent Fresnel factor, and roughness-guided filtering approximates glossy reflection.Only the intrinsic reflectance enters texture space; the angle-dependent factor is evaluated during reflection rendering.Source: §A.1 · p. 13
• Texture-space inpainting aggregates view-independent intrinsic attributes rather than the view-dependent glossy component, avoiding direct combination of highlights and reflections across views.The glossy component is evaluated from the edited scene.Source: §A.1 · p. 13

A.2 TEXTURE-SPACE INPAINTING DETAILS

• A fitted plane provides the shared geometric reference for texture projection, object masks, Gaussian initialization, and surface regularization.The plane is estimated by minimizing background Gaussian centers’ squared distances to it.Source: §A.2 · p. 13
• Plane coordinates use two tangent axes and a regular texel grid; texture coordinates and grid coordinates serve distinct geometric and sampling roles.A stride of sT texels corresponds to distances sT ∆u and sT ∆v on the surface.Source: §A.2 · pp. 13–14
• Camera rays intersect the plane to produce texture coordinates; aggregation requires a forward, in-domain correspondence and a visible background observation.A ray must not be parallel to the plane, and retained objects must not occlude the observation.Source: §A.2 · p. 14
• Intrinsic observations are bilinearly splatted to neighboring texels and weighted by the absolute dot product of the plane normal and viewing direction.Grazing views receive lower weight; the accumulated support denominator is stabilized by ϵ, which does not make unobserved texels observed.Source: §A.2 · p. 14
• The completion mask is the union of selected objects’ projected footprints, distinct from the observation-validity set that determines which values enter the texture.Object centers closest to the fitted plane are projected to form their footprints.Source: §A.2 · p. 14
• After object deletion, the method aggregates diffuse, Fresnel, and roughness textures, then inpaints all three with the same footprint mask.Generated values are retained inside the mask, aggregated inputs remain unchanged outside it, and normals and the scene-dependent glossy term are not inpainted.Source: §A.2 · p. 15
• Completed intrinsic maps are reprojected to camera views, where pixels at the same plane location receive the same completed texture value.Targets must be in front of the camera, inside the texture domain, and unoccluded by retained content; they may cover regions without original background observations.Source: §A.2 · p. 15

A.3 GAUSSIAN INITIALIZATION AND OPTIMIZATION DETAILS

• Completion Gaussians are initialized within the texture-inpainting mask, using sampled texture locations to set their initial centers and appearance.Their initial scales are determined by the sampling spacing.Source: §A.3 · p. 15
• The initialization mask matches the texture-space inpainting mask, linking the region sampled for completion Gaussians to the region receiving texture supervision.The objective combines intrinsic supervision with the three weighted geometry penalties.Source: §A.3 · pp. 15–16
• Soft surface and anchor penalties discourage normal and tangential center drift, while the scale penalty discourages expansion or collapse relative to the sampling footprint.These penalties permit local adjustments and do not prescribe exact optimized centers; the surface term does not directly constrain orientation.Source: §A.3 · p. 16
• Only the completion Gaussians are optimized to fit fixed reprojected intrinsic-map targets, using a masked L1 loss and geometry penalties.The retained Gaussians, completed textures, targets, and supervision masks remain fixed.Source: §A.3 · p. 16

B SUPPLEMENTARY RESULTS

• The supplementary comparisons complement the main-text texture-space inpainting ablation.The cited passage identifies Figure 5 and Table 2 as the main-text comparisons.Source: B SUPPLEMENTARY RESULTS · p. 17
• Figure B.1 presents a visual comparison of multi-object removal.The caption does not state a specific outcome.Source: Figure B.1 · p. 17
• Figure B.2 compares screen-space and texture-space inpainting for both single-object and multi-object removal.These comparisons supplement the texture-space inpainting ablation in Figure 5 and Table 2.Source: Figure B.2 · p. 17, B SUPPLEMENTARY RESULTS · p. 17