PaperlayerView exact revision on arXiv

Source-linked AI summary

Efficient Geometry-aware 3D Generative Adversarial Networks

Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, Gordon Wetzstein

arXiv:2112.07945v2cs.CVcs.AIcs.GRcs.LG

TL;DR

Existing 3D GANs lag in image quality, resolution, and 3D reconstruction quality, while unsupervised generation of multi-view-consistent images and shapes from single-view photographs remains challenging. This work introduces an efficient hybrid 3D GAN and demonstrates state-of-the-art 3D-aware synthesis on FFHQ and AFHQ Cats, alongside high-quality 3D geometry.

  • Problem

    Existing 3D GANs lag in image quality, resolution, and 3D reconstruction quality, with computational inefficiency identified as a primary reason.

    Source: §1 · p. 1

  • Method

    The framework combines an efficient, expressive tri-plane hybrid explicit–implicit 3D GAN with pose conditioning and dual discrimination.

    Source: §1 · p. 2

  • Results

    The approach demonstrates state-of-the-art unconditional 3D-aware synthesis on FFHQ and AFHQ Cats and produces high-quality 3D geometry learned from in-the-wild 2D images.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The approach takes significant steps toward photorealistic 3D-aware image synthesis and high-quality unsupervised shape generation.

    Source: §6 · p. 8

    The method requires the dataset’s camera-pose distribution, and pose conditioning does not fully disentangle appearance from pose, leaving geometry ambiguities unresolved.

    Source: §6 · p. 8

Abstract

from arXiv · show

Unsupervised generation of high-quality multi-view-consistent images and 3D shapes using only collections of single-view 2D photographs has been a long-standing challenge. Existing 3D GANs are either compute-intensive or make approximations that are not 3D-consistent; the former limits quality and resolution of the generated images and the latter adversely affects multi-view consistency and shape quality. In this work, we improve the computational efficiency and image quality of 3D GANs without overly relying on these approximations. We introduce an expressive hybrid explicit-implicit network architecture that, together with other design choices, synthesizes not only high-resolution multi-view-consistent images in real time but also produces high-quality 3D geometry. By decoupling feature generation and neural rendering, our framework is able to leverage state-of-the-art 2D CNN generators, such as StyleGAN2, and inherit their efficiency and expressiveness. We demonstrate state-of-the-art 3D-aware synthesis with FFHQ and AFHQ Cats, among other experiments.

1. Introduction

The introduction argues that 3D GANs lag 2D models in image quality and geometry because existing 3D representations and rendering are inefficient, while image-space upsampling sacrifices consistency. It proposes an efficient, hybrid 3D-aware framework for multi-view-consistent synthesis and geometry learning from single-view 2D images.

1. Introduction

  • Existing explicit and implicit 3D representations are too memory-intensive or slow for high-resolution GAN training, while image-space upsampling sacrifices view consistency and geometry quality.Training requires rendering tens of millions of images, making high-resolution neural volume rendering computationally infeasible with these representations.Source: §1 · pp. 1–2
  • The tri-plane-based hybrid explicit–implicit representation aims to improve rendering speed and memory efficiency without sacrificing expressiveness, enabling high-resolution geometry-aware image synthesis.The framework is designed to avoid the computational constraints that limited prior rendering resolutions and drove reliance on image-space convolutional upsampling.Source: §1 · p. 2
  • Dual discrimination regularizes image-space approximations, while pose conditioning supports multi-view consistency and models pose-correlated attributes such as facial expressions.The generator is intended to decouple pose-correlated attributes at inference while faithfully modeling their joint distributions in training data.Source: §1 · p. 2
  • Decoupling feature generation from neural rendering lets the framework leverage 2D CNN generators such as StyleGAN2 while retaining 3D multi-view-consistent neural volume rendering.Source: §1 · p. 2
  • The authors report state-of-the-art 3D-aware image synthesis on FFHQ and AFHQ Cats, alongside high-quality 3D geometry learned from in-the-wild 2D images.The framework learns from collections of single-view photographs without geometry supervision or multi-view image collections.Source: §1 · p. 2

2. Related work

Existing 3D-aware approaches face tradeoffs in memory, computation, or view consistency. The paper positions its tri-plane explicit–implicit representation as an efficient, 3D-centered alternative.

Neural scene representation and rendering

  • Explicit voxel grids evaluate quickly but use substantial memory, while implicit coordinate networks [43] [45] [52] [60] [66] save memory and handle complex scenes but are slow to query.The two representation types therefore offer complementary benefits.Source: §2 · p. 2
  • The tri-plane hybrid stores features on axis-aligned planes and aggregates them with a lightweight implicit decoder for efficient rendering, with training times comparable to modern 2D GANs.It resembles earlier plane-based hybrids [53] but uses a distinct design.Source: §2 · p. 2

Generative 3D-aware image synthesis

  • Mesh methods lack high-fidelity expressiveness [64], voxel GANs face memory and 3D-convolution costs [47] [48] [68], and fully implicit generators are slow to query [58].2D upsampling can lack view consistency without a 3D bias [49], while sparse block volumes suit mostly empty scenes and generalize poorly.Source: §2 · p. 3
  • The framework uses StyleGAN2, efficient tri-plane volume rendering, and dual discrimination after 2D upsampling [49] to avoid view inconsistencies, including at steep angles and in 360° viewing conditions.Unlike StyleGAN2-based 2.5D GANs that generate images and depth maps [59], it works naturally across these viewing conditions.Source: §2 · p. 3
  • Against StyleNeRF and CIPS-3D [73], which operate primarily in image-space, the 3D-focused method reports greater view consistency, high-quality shapes, and superior FID on FFHQ and AFHQ.The passages give no FID values.Source: §2 · p. 3

3. Tri-plane hybrid 3D representation

This section introduces a hybrid explicit–implicit tri-plane representation that combines expressive plane features with a lightweight decoder for efficient 3D scene representation.

3. Tri-plane hybrid 3D representation

  • Tri-planes outperform dense feature volumes and fully implicit representations on PSNR and SSIM, and are 3–8× faster than a fully implicit Mip-NeRF network.The single-scene overfitting experiment fit 360° Tanks & Temples views; neither voxels nor tri-planes model view-dependent effects.Source: Figure 3 · p. 3, §3 · pp. 3–4
  • Each 3D position is projected onto three orthogonal feature planes; bilinearly interpolated features are summed, then a small MLP decodes color and density for neural volume rendering.The planes have resolution N × N × C, with N denoting spatial resolution and C the channel count.Source: §3 · p. 3
  • Keeping the decoder small and shifting expressive power into explicit features reduces neural-rendering computation versus fully implicit MLP architectures [45] without losing expressiveness.The single-scene overfitting experiment used a larger decoder than the GAN experiments to prioritize expressiveness over speed.Source: §3 · pp. 3–4
  • Tri-planes scale as O(N^2), rather than dense voxels’ O(N^3), enabling higher-resolution features at equal capacity and memory; their planes can also be generated by an off-the-shelf 2D CNN.The experiment used N = 512, C = 48 planes and a four-layer MLP with 128 hidden units per layer and Fourier feature encoding [66].Source: §3 · p. 4

4. 3D GAN framework

The framework trains a geometry-aware 3D GAN from single-view photographs without explicit 3D or multiview supervision, combining generated tri-plane features, neural volume rendering, and image-space refinement. Camera conditioning and dual discrimination help model dataset pose correlations while supporting multi-view consistency.

4. 3D GAN framework

  • The 3D GAN learns from single-view photographs without explicit 3D or multiview supervision, using detected camera intrinsics and extrinsics.It trains end-to-end from random initialization with a non-saturating GAN loss and R1 regularization, first at 64² neural-rendering resolution and then briefly at 128²; density smoothness regularization helped reduce shape artifacts.Source: §4 · p. 4

4.1. CNN generator backbone and rendering

  • StyleGAN2 generates 32-channel tri-planes, whose aggregated features yield density and 32-channel samples for neural volume rendering into feature images.The backbone outputs a 256 × 256 × 96 feature image, reshaped into three planes; a one-hidden-layer, 64-unit MLP decoder predicts density and features without positional encoding or coordinate or view-direction inputs.Source: §4.1 · pp. 4–5

4.2. Super resolution

  • Moderate-resolution neural renderings are upsampled to 256² or 512² final images because high-resolution volume rendering remains too slow for interactive framerates.Two StyleGAN2-modulated convolution blocks refine the 32-channel feature image; per-pixel noise inputs are disabled to reduce texture sticking.Source: §4.2 · p. 5

4.3. Dual discrimination

  • Dual discrimination checks low-resolution rendered RGB alongside the super-resolved image, encouraging both realistic outputs and consistency between resolutions.The discriminator receives the upsampled first three feature channels concatenated with the final image, and camera intrinsics and extrinsics as conditioning to guide 3D priors.Source: §4.3 · p. 5, Figure 5 · p. 5

4.4. Modeling pose-correlated attributes

  • Generator pose conditioning models pose-correlated dataset biases during training, then uses a fixed conditioning pose at inference to prevent scene shifts as the camera moves.Because conditioning on the rendering pose can yield camera-facing 2D billboards, training randomly swaps in another conditioning pose with 50% probability.Source: §4.4 · pp. 5–6

5. Experiments and results

Experiments evaluate unconditional 3D-aware generation on FFHQ and AFHQv2 Cats, finding improved image quality alongside strong view consistency and geometry quality. Ablations and applications examine expression consistency, style mixing, and single-view 3D reconstruction.

5. Experiments and results

  • Experiments use FFHQ and AFHQv2 Cats, augment both datasets with horizontal flips, and apply pose estimators to obtain approximate camera extrinsics.AFHQv2 methods use transfer learning from corresponding FFHQ checkpoints; the method additionally uses adaptive data augmentation for AFHQv2 512².Source: §5 · p. 6

5.1. Comparisons

  • Across FFHQ and AFHQ Cats, the method significantly improves FID toward StyleGAN2 512² while maintaining state-of-the-art view consistency, geometry quality, and pose accuracy.Evaluation uses FID, identity consistency, depth accuracy, and pose accuracy; the supplied passages do not provide Table 2’s individual metric values.Source: Table 2 · p. 7, §5.1 · p. 7
  • Compared with π-GAN, GIRAFFE [49], and Lifting StyleGAN [59], the method produces higher-quality, more view-consistent images and higher-fidelity 3D geometry.The comparisons highlight detailed glasses and hair strands; GIRAFFE’s view-inconsistent convolutions cause poor shapes and identity shifts, while π-GAN and Lifting StyleGAN struggle with photorealism and detailed shapes.Source: §5.1 · p. 7
  • At 512² final resolution with 128² neural rendering, the method achieves real-time rendering and is significantly faster than fully implicit π-GAN.Caching tri-plane features for consecutive frames of a static scene further improves speed; the method remains slower than Lifting StyleGAN [59] and GIRAFFE [49].Source: §5.1 · p. 7, Table 3 · p. 7

5.2. Ablation study

  • Dual discrimination reduces smile-coefficient variation across video frames but slightly worsens FID; generator pose conditioning enables both the best FID and view-consistent results.Pose conditioning lets the generator model pose-correlated attributes while decoupling them at inference; smile variation is measured using FACS coefficients from videos.Source: §5.2 · p. 7, Table 4 · p. 7

5.3. Applications

  • The StyleGAN2-based 3D representation inherits latent-space properties that enable semantic image manipulation through style mixing.Figure 8 shows style-mixing results on FFHQ at 512².Source: §5.3 · p. 8, Figure 8 · p. 8
  • Pivotal tuning inversion fits test images, and the learned FFHQ 3D prior enables high-quality geometry recovery from a single view.The paper identifies few-shot 3D reconstruction and novel-view synthesis as future directions.Source: §5.3 · p. 8

6. Discussion

The discussion presents the framework as a step toward photorealistic 3D-aware image synthesis and high-quality unsupervised shape generation, while identifying limitations, risks, and directions for future work.

Limitations and future work

  • Generated shapes may still contain artifacts and lack fine details such as individual teeth; stronger geometry priors or radiance-field density regularization are proposed directions [51] [67] [69].Source: §6 · p. 8
  • The model requires the dataset’s camera-pose distribution; learning it on the fly has been proposed [49] but can diverge, and pose conditioning does not fully disentangle appearance from pose.Geometry ambiguities remain: concave eye sockets can create the illusion of eyes following the camera despite view-consistent renderings that reflect the underlying geometry.Source: §6 · p. 8
  • Beyond StyleGAN 2, image-to-image translation or Transformer-based backbones may enable conditional-synthesis applications.Source: §6 · p. 8

Ethical considerations

  • Single-view reconstruction and style mixing could be misused to edit real people’s imagery, threatening society through misinformation or reputational harm; face results may also reflect dataset biases.Source: §6 · p. 8

Conclusion

  • Combining an efficient explicit–implicit representation, pose-aware convolutional generator, and dual discriminator advances photorealistic 3D-aware synthesis and unsupervised shape generation.The authors suggest this could support rapid 3D prototyping, more controllable image synthesis, and shape reconstruction from temporal data.Source: §6 · p. 8

Supplemental Material Efficient Geometry-aware 3D Generative Adversarial Networks

The supplement outlines additional experiments and visual results, implementation and experiment details, and artifacts identified as possible targets for future work.

Additional experiments and visual results

  • It presents additional experiments and visual results, including a supplemental video with a live demonstration of real-time synthesis.The video also contains additional visual results.Source: Supplemental Material Efficient… · p. 12

Implementation

Experiment details and artifacts

  • Experiment details address datasets, baselines, and further explanations of experiments such as inversion; the supplement also discusses artifacts as possible future-work targets.Source: Supplemental Material Efficient… · p. 12

1. Additional experiments

Additional experiments examine pose-linked expression warping and test the method’s multi-view consistency, conditioning robustness, steep-view extrapolation, and performance across datasets.

1.1. Analyzing pose/facial expression correlation in FFHQ

  • FFHQ subjects facing the camera are more likely to smile than those facing away, creating a pose–expression correlation that can incentivize expression warping as the camera moves.Dual discrimination and generator pose conditioning are proposed to reduce this effect.Source: §1.1 · p. 12

1.2. COLMAP reconstruction

  • COLMAP reconstructed a dense, well-defined point cloud from 128 synthesized video frames on an oval camera trajectory, supporting highly multi-view-consistent renderings.The reconstruction used COLMAP’s automatic mode without specified camera parameters.Source: §1.2 · pp. 12–13, Figure 2 · p. 12

1.3. Regularizing generator pose conditioning

  • Without pose swapping, generator pose conditioning can produce camera-facing billboards; randomly swapping conditioning poses prevents this degenerate solution.The swapping probability starts at 100%, decays to 50% over the first 1M images, and remains at 50% thereafter.Source: §1.3 · p. 13, Figure 3 · p. 13

1.4. Robustness to imprecise camera poses

  • Without discriminator pose conditioning, FFHQ models flatten heads onto planes; 3σ pose noise still yields accurate 3D shapes, while at 4σ some scenes flatten.The noise levels are standard deviations of Gaussian corruption applied to camera poses.Source: §1.4 · pp. 13–14, Figure 4 · p. 13

1.5. Extrapolation to steep camera angles

  • Our 3D scaffold extrapolates more reasonably to rare yaw and pitch views than GIRAFFE, while π-GAN degrades at pose-distribution edges and lifted StyleGAN shows inaccurate geometry.FFHQ contains few extreme-yaw views and even fewer extreme-pitch views; GIRAFFE also produces degraded, view-inconsistent images when forced beyond its sampled training poses.Source: §1.5 · p. 14

1.6. Additional quantitative results

  • Strong relative performance on uniformly posed ShapeNet Cars suggests the method is not limited to face-forward FFHQ and AFHQv2 datasets.The expanded quantitative evaluation includes KID for all datasets and image-quality metrics for Cars.Source: §1.6 · p. 14, Table 1 · p. 14

2. Additional visual results

The additional visual results illustrate shape-level style mixing, single-image face reconstruction, and synthesis across diverse datasets and latent codes. They show how layer modulations affect shape detail, reconstructions can support photo-to-avatar applications, and latent interpolations remain smooth in renderings and shapes.

Style mixing, in shapes

  • Style mixing inherits most shape structure from low-resolution layers, while high-resolution layers influence details such as eyes and hair.These results use a model trained without style-mixing regularization; the fine input also controls rendered-image colors.Source: §2 · p. 14, Figure 6 · p. 16

Additional single image 3D reconstructions

  • PTI produces additional single-image 3D reconstructions of FFHQ 512² test faces, suggesting applications such as photo-to-avatar creation.The pipeline does not require explicit 3D ground-truth training data; the examples demonstrate use of the generator’s learned prior over facial features.Source: §2 · pp. 14, 16, Figure 7 · p. 16

Shapenet Cars

  • ShapeNet Cars renderings demonstrate operation on datasets spanning the full 360° camera azimuth and 180° elevation, unlike 2.5D GANs intended for face-forward datasets.The examples are uncurated and rendered from random camera poses.Source: §2 · p. 16

Additional selected examples synthesized with AFHQv2 Cats

  • Selected AFHQv2 Cats examples show renderings and shapes synthesized by the method trained at 512² resolution.The examples are from the AFHQv2 Cats dataset.Source: §2 · p. 16

Uncurated examples synthesized with AFHQv2 Cats

  • Uncurated AFHQv2 Cats examples compare GIRAFFE, π-GAN, and the method at 256², 128², and 512² image resolutions, respectively.Source: §2 · p. 16

Uncurated examples synthesized with FFHQ

  • Uncurated FFHQ face examples are synthesized at 512² resolution using truncation with ψ = 0.5.Source: §2 · p. 16

Latent code interpolation

  • Linear latent-code interpolations show smooth changes in both color renderings and underlying shapes, reflecting the StyleGAN2 backbone’s well-behaved latent space.The examples use selected codes from a model trained on FFHQ 512².Source: §2 · p. 16

Additional selected examples synthesized with FFHQ

  • Additional selected FFHQ examples show renderings and shapes synthesized by the method trained at 512² resolution.Source: §2 · p. 16

3. Implementation details

The implementation uses StyleGAN2 components and two-stage neural-rendering training, and reports training throughput alongside an inference-time sampling tradeoff.

Two-stage training

  • Two-stage training uses 64^2 rendering for 25M images, then 128^2 for 2.5M; it provides roughly 2× speed-up over full-resolution-from-scratch training with similar results.Rendering resolution is gradually increased pixel-by-pixel over 1M images, then held at 128^2 for 1.5M fine-tuning images.Source: §3 · pp. 16, 17, 20

Backbone

  • The StyleGAN2 backbone runs at 256^2 and outputs 96 channels reshaped into three 256×256×32 planes; the pipeline is trained end-to-end without pretrained checkpoints.The decoder is a one-hidden-layer MLP with 64 units and softplus activation, mapping 32-channel features to density plus 32 features; super-resolution uses two modulated-convolution blocks of depths 128 and 64.Source: §3 · pp. 17–18

Dataset-specific training

  • Large datasets such as FFHQ and ShapeNet Cars are trained from scratch, while AFHQv2 fine-tunes FFHQ-trained models; ADA is used only for the method at 512^2.For AFHQv2, training proceeds for 6.2M images at 64^2 and 2.6M more while fine-tuning rendering resolution to 128^2.Source: §3 · pp. 18, 20, 21

Discriminator

  • The StyleGAN2 discriminator accepts six-channel inputs for dual discrimination and is conditioned on camera parameters to help prevent degenerate shapes.Source: §3 · p. 19

Density Regularization

  • Density regularization minimizes the L1 difference between densities at randomly sampled volume points and Gaussian-perturbed points, encouraging a smoother field and reducing seams and shape artifacts.It is applied to 1,000 point pairs every four training iterations.Source: §3 · pp. 19–20

Training

  • Training uses batch size 32, learning rates 0.002 for the discriminator and 0.0025 for the generator, and mixed precision only in high-resolution discriminator and super-resolution blocks.R1 regularization uses γ = 1 except on ShapeNet Cars, where γ = 0.1; training omits style-mixing regularization and gradually reduces input blur over the first 200K images.Source: §3 · pp. 19–20
  • Training ran at 24 s/kimg at 64^2 and 46 s/kimg at 128^2 on 8 Tesla V100 GPUs, versus StyleGAN3-R’s 20 s/kimg; total time was 8.5 days versus 6 days.The reported totals compare this framework with StyleGAN3-R on similar hardware.Source: §3 · p. 20

Inference-time depth samples

  • For thin-surface videos, doubling coarse and fine samples from 48 to 96 reduces flicker but lowers frame rates from 36 to approximately 24 fps with tri-plane caching.Main-manuscript images use the default 48 samples; FFHQ and AFHQv2 use 48 uniform plus 48 importance samples per ray, while ShapeNet Cars uses 64 of each.Source: §3 · pp. 18, 20

4. Experiment details

The experiments specify baseline implementations, dataset preparation, inversion and evaluation procedures, and geometry visualization. The setup spans FFHQ, AFHQv2 Cats, ShapeNet Cars, and a single-scene Tanks & Temples comparison.

4.1. Baselines

  • Baselines used official π-GAN and GIRAFFE code trained to convergence, while Lifting StyleGAN used its official tighter-crop pretrained model and reported FID.The Lifting StyleGAN model was used because training from scratch on the less-cropped FFHQ data did not match its pretrained model’s quality.Source: §4.1 · p. 21
  • StyleGAN2 tuning compared config F and the 512 × 512 config from, sweeping R1 regularization; config F gave the best AFHQv2 result after 10M images at γ = 1.The sweep tested γ = {0.2, 0.5, 1, 2, 5, 10, 20}.Source: §4.1 · pp. 21–22

4.2. Dataset Details

  • FFHQ preparation detected and posed uncropped images, cropped faces, assumed fixed intrinsics, and retained 69,957 images; AFHQv2 experiments used its approximately 5,000-image Cats split.FFHQ used a focal length of 4.26 × image width; AFHQv2 used the same intrinsics, landmark detection and Perspective-n-Point [3] for poses, and horizontal flips.Source: §4.2 · pp. 22–23
  • ShapeNet Cars validation used 2,457 synthetic cars rendered at 128² resolution from 50 sphere-sampled views each, with known camera parameters and no image-space augmentation.The setup followed and evaluated views spanning all angles.Source: §4.2 · p. 23

4.3. Single scene overfitting.

  • The Tanks & Temples Family-scene comparison tested tri-planes, a parameter-matched voxel hybrid, and Mip-NeRF using the preprocessing and split from.Both hybrid representations used two-layer, 128-hidden-unit decoders with Fourier features; they trained for 200K iterations, versus 1M for Mip-NeRF, with 512 depth and 256 importance samples per ray and ray batches of 6,400.Source: §4.3 · p. 24

4.4. Pivotal tuning inversion.

  • Pivotal tuning inversion optimized each cropped, 512² test image’s latent code for 500 iterations, then fine-tuned generator weights for 500 more.For grayscale inversion, RGB renders were converted to luminance using Y = 0.299R + 0.587G + 0.114B; optimization used 400 latent steps and 250 fine-tuning steps.Source: §4.4 · p. 24

4.5. Evaluation Metrics

  • Evaluation measured FID and KID on 50k generated images against all training images, plus geometry depth L2 error against pseudo-ground-truth maps.Geometry used 1,024 random-pose images, face-region depth maps, and zero-mean, unit-variance normalization; FID and KID used the StyleGAN3 implementation.Source: §4.5 · p. 24
  • Other evaluations measured FFHQ multi-view identity similarity with ArcFace, pose drift against detected pitch, yaw, and roll, runtime over 400 batch-size-one frames, and facial-expression variability.FACS estimation compared the ‘mouthSmile L’ and ‘mouthSmile R’ coefficients across videos using a proprietary tracker measuring 53 blendshape coefficients.Source: §4.5 · p. 24

5. Discussion

The method improves geometry over previous methods but retains artifacts tied to shape illusions, eyeglasses, seams, and its single-representation treatment of backgrounds.

5.1. Shape artifacts

  • Despite improved geometry over previous methods, artifacts persist: sunken sockets and mouth creases can create camera-following eyes or view-inconsistent effects faithful to underlying shapes.Stronger dataset priors, such as convex eyeballs, may help resolve these artifacts.Source: §5.1 · pp. 24–25
  • The method produces more-detailed eyeglasses than previous methods but tends to make their sides opaque like “goggles,” where empty space should be.Neural rendering that models lens refraction accurately may enable more faithful reconstruction of eyeglasses and other transparent objects.Source: §5.1 · p. 25
  • Some generated shapes show a face-to-head seam; optional density regularization helps reduce it, while hybrid-SDF rendering solutions may improve geometry.The cited hybrid-SDF solutions have shown promising results in robust geometry recovery from images.Source: §5.1 · p. 25
  • Because the scene uses a single 3D representation without explicit background handling, backgrounds become textured surfaces fused to foreground objects.Separate 3D background representations may enable foreground isolation.Source: §5.1 · p. 25

Related papers

Loading 2112.07945v2…