ndeda/neural-object-reconstruction-abo

Neural Object Reconstruction and Relighting Research Dataset

Uncertainty-Aware Multi-View Material Decomposition for Relightable 3D Objects

Project Description

This repository contains the data infrastructure and research materials for a study on 3D object reconstruction, material decomposition and relighting from multi-view RGB observations.

The project investigates whether an object captured using a short smartphone video can be converted into a structured 3D representation that preserves useful information about geometry, appearance, material properties and prediction uncertainty.

The work combines several areas of computer vision and computer graphics, including:

• multi-view reconstruction;
• object segmentation;
• monocular depth estimation;
• camera pose estimation;
• 3D Gaussian Splatting;
• neural rendering;
• inverse rendering;
• material decomposition;
• uncertainty estimation;
• differentiable rendering;
• novel-view synthesis;
• relighting.

The research is designed as a modular pipeline. Individual components are evaluated independently before being integrated into the complete system.

The central objective is to move beyond appearance-only reconstruction toward a representation that can be used for controlled rendering under different lighting conditions.

Research Question

The primary research question is:

Can uncertainty-aware multi-view material decomposition improve the reliability and controllability of 3D Gaussian representations reconstructed from casual smartphone video?

Research Hypothesis

The working hypothesis is that explicitly estimating intrinsic material properties together with their uncertainty can produce a more useful representation for novel-view rendering and relighting than a conventional RGB-only 3D representation.

This hypothesis will be evaluated through controlled experiments, baseline comparisons and ablation studies.

The hypothesis should not be considered experimentally validated until the corresponding experiments have been completed.

Research Objectives

The project has the following objectives:

1. Develop an object segmentation pipeline for casual video.
2. Investigate temporal segmentation and high-resolution foreground refinement.
3. Estimate depth from RGB observations.
4. Estimate camera intrinsics and poses.
5. Establish a 3D Gaussian Splatting reconstruction baseline.
6. Build a controlled synthetic multi-view dataset.
7. Develop multi-view material decomposition.
8. Estimate albedo, roughness, specular properties and surface normals.
9. Estimate illumination separately from intrinsic material properties.
10. Estimate uncertainty associated with material predictions.
11. Integrate material information into a 3D Gaussian representation.
12. Develop material-aware differentiable rendering.
13. Evaluate novel-view synthesis.
14. Evaluate relighting under unseen illumination.
15. Investigate synthetic-to-real transfer.
16. Evaluate the complete system using real smartphone captures.
17. Analyze failure cases and uncertainty.

Dataset Source

Amazon Berkeley Objects

The primary source of 3D object assets for this project is the Amazon Berkeley Objects (ABO) dataset.

The project uses selected 3D assets rather than attempting to reproduce the complete original dataset.

The 3D assets are accessed through Amazon S3 as part of the data preparation pipeline.

The resulting collection is a research-derived subset prepared for controlled experiments and synthetic data generation.

The project does not claim ownership of the original ABO assets.

Users of ABO-derived material should consult the original dataset documentation, licensing terms and citation requirements.

Why Amazon Berkeley Objects Is Used

The research requires actual 3D object assets because material decomposition and relighting experiments require controlled knowledge of the underlying object.

A 3D object allows the project to generate multiple observations while controlling variables such as:

• camera position;
• camera orientation;
• object rotation;
• illumination;
• material properties;
• background;
• rendering configuration.

This provides a controlled environment for generating supervision that is difficult to obtain from ordinary photographs.

Amazon S3 is used as the storage and retrieval layer for the relevant 3D assets.

The current data workflow therefore consists of:

1. metadata discovery;
2. object selection;
3. candidate filtering;
4. S3 asset retrieval;
5. GLB validation;
6. dataset organization;
7. manifest generation.

Synthetic Dataset Generation

Selected ABO 3D assets are used as source objects for controlled synthetic rendering.

The synthetic data generation process is intended to produce multiple views of the same object while varying camera, material and lighting conditions.

The intended process is:

ABO 3D Asset | v Object Selection | v Quality Filtering | v Controlled Rendering | +---- Camera Variation +---- Object Rotation +---- Lighting Variation +---- Material Variation +---- Background Variation | v Multi-View Training / Evaluation Data

Depending on the rendering configuration, generated samples can contain synchronized channels such as:

• RGB;
• depth;
• segmentation;
• surface normals;
• albedo;
• roughness;
• specular properties;
• illumination;
• camera parameters.

The synthetic dataset is intended to provide controlled supervision for material decomposition, uncertainty estimation and relighting.

Data Roles

Data Source	Role
Amazon Berkeley Objects	Primary 3D object source
Amazon S3	3D asset retrieval and storage
Project synthetic dataset	Controlled training and evaluation
OpenIllumination	Material and illumination evaluation
OpenRooms	Material and lighting research
Tanks and Temples	Multi-view reconstruction evaluation
Project smartphone captures	Real-world evaluation

The datasets listed above serve different purposes. The project does not assume that a single dataset is sufficient for every research question.

Research Pipeline

The planned end-to-end system consists of the following stages:

Smartphone Video
       |
       v
Frame Extraction
       |
       v
Object Segmentation
       |
       v
Mask Refinement
       |
       v
Depth Estimation
       |
       v
Camera Pose Estimation
       |
       v
3D Gaussian Splatting
       |
       v
Multi-View Feature Learning
       |
       v
Material Decomposition
       |
       v
Uncertainty Estimation
       |
       v
Material-Aware Gaussian Representation
       |
       v
Differentiable Rendering
       |
       v
Novel Lighting
       |
       v
Relightable 3D Object

The individual stages are developed and evaluated independently before being combined into the complete research system.

The segmentation stage investigates SAM 2 and BiRefNet.

Camera estimation uses COLMAP as the initial structure-from-motion baseline.

3D Gaussian Splatting provides the initial reconstruction representation.

The material stage investigates albedo, roughness, specular properties, surface normals and illumination.

Uncertainty estimation is used to study the reliability of material predictions.

The final rendering stage investigates whether the resulting representation can support novel-view synthesis and relighting.

Object Segmentation

Object segmentation separates the target object from the surrounding scene before reconstruction.

The project investigates temporal segmentation together with high-resolution foreground refinement.

SAM 2

SAM 2 is investigated for:

• object segmentation
• temporal propagation
• video tracking
• segmentation confidence

BiRefNet

BiRefNet is investigated for:

• foreground refinement
• boundary refinement
• high-resolution masks

The research also considers a fusion stage combining information from both approaches.

Segmentation quality is evaluated independently because errors at this stage can propagate into depth estimation, reconstruction and material decomposition.

Depth Estimation

Depth estimation provides geometric information for downstream reconstruction.

The initial system uses pretrained depth estimation models.

The general process is:

RGB Frame
    |
    v
Depth Model
    |
    v
Depth Map

Synthetic data may later be used to investigate domain-specific adaptation.

Camera Pose Estimation

Camera estimation provides the poses required for multi-view reconstruction.

The initial reconstruction workflow uses COLMAP for structure-from-motion.

The resulting information includes:

• camera intrinsics
• camera extrinsics
• feature correspondences
• estimated camera poses
• sparse reconstruction information

Accurate camera estimation is important because pose errors can affect 3D reconstruction and multi-view material consistency.

3D Gaussian Splatting

3D Gaussian Splatting is used as the primary reconstruction baseline.

A conventional Gaussian representation contains parameters associated with:

• position
• scale or covariance
• rotation
• opacity
• appearance

The research investigates extending this representation with material information.

A material-aware Gaussian may contain:

• position
• scale
• rotation
• opacity
• albedo
• roughness
• specular properties
• normal information
• uncertainty

The purpose is to investigate whether explicit material information can improve rendering and relighting compared with an RGB-only representation.

Multi-View Material Decomposition

Material decomposition is the central research component.

Rather than predicting material properties independently from each image, the project investigates using multiple observations of the same object.

The target properties include:

Albedo

An estimate of intrinsic surface appearance.

Roughness

An estimate of the surface roughness response.

Specular Properties

Information describing the specular component of surface appearance.

Surface Normals

Surface orientation information useful for geometry and rendering.

Illumination

An estimate of lighting contributing to the observed appearance.

The objective is not to claim perfect physical recovery of material properties.

The research evaluates whether the learned representation is useful for reconstruction and relighting.

Uncertainty Estimation

The project treats uncertainty as an important part of the material prediction problem.

Different surface regions have different levels of observability.

Potentially difficult regions include:

• reflective surfaces
• transparent objects
• dark surfaces
• occluded regions
• unseen surfaces
• motion-blurred observations
• poorly reconstructed regions

The research investigates whether uncertainty estimates can identify areas where material predictions are less reliable.

Planned analysis includes:

• uncertainty/error correlation
• calibration
• confidence maps
• high-error/high-uncertainty overlap

The relationship between uncertainty and prediction error will be measured experimentally rather than assumed.

Relighting

Relighting is the primary downstream application of the proposed representation.

The intended experiment is:

1. capture an object
2. reconstruct its geometry
3. estimate its material properties
4. construct the material-aware representation
5. change the lighting conditions
6. render the reconstructed object under the new lighting

The evaluation investigates whether the rendered appearance remains consistent with the estimated material properties.

This distinguishes the research from systems that primarily reproduce the observed RGB appearance.

Experimental Design

The research uses progressively stronger experimental configurations.

Experiment 1 — RGB Gaussian Baseline

A conventional RGB-based 3D Gaussian Splatting system.

Purpose: establish the baseline quality of novel-view reconstruction.

Experiment 2 — RGB + Depth

Depth information is incorporated into reconstruction.

Purpose: measure the contribution of explicit geometric information.

Experiment 3 — Single-View Material Prediction

Material properties are predicted from individual observations.

Purpose: establish a single-view material baseline.

Experiment 4 — Multi-View Material Prediction

Multiple observations of the same object are processed jointly.

Purpose: measure the contribution of multi-view information.

Experiment 5 — Material-Aware Gaussian Representation

Material properties are stored within the Gaussian representation.

Purpose: evaluate the effect of explicit material representation on rendering and relighting.

Experiment 6 — Uncertainty-Aware Representation

Uncertainty estimates are incorporated into material prediction and the 3D representation.

Purpose: determine whether uncertainty provides useful information about prediction reliability.

Experiment 7 — Complete Pipeline

The complete proposed system combines:

• segmentation
• depth
• camera estimation
• 3D reconstruction
• material decomposition
• uncertainty
• material-aware Gaussians
• differentiable rendering
• relighting

Ablation Studies

Component-level ablations are used to determine which parts of the system contribute to final performance.

Planned ablations include:

• without depth
• without multi-view information
• without material features
• without uncertainty
• without segmentation refinement

The segmentation pipeline may also be compared using:

SAM 2
   vs.
BiRefNet
   vs.
SAM 2 + BiRefNet
   vs.
SAM 2 + BiRefNet + Learned Fusion

The purpose is to separate the contribution of individual components from the performance of the complete system.

Evaluation Protocol

The project evaluates different components using task-specific metrics.

Segmentation

• Intersection over Union (IoU)
• Dice / F1
• Boundary F-score

Depth

• RMSE
• Absolute Relative Error
• threshold accuracy

Novel-View Reconstruction

• PSNR
• SSIM
• LPIPS
• geometric error
• multi-view consistency

Material

Albedo:

• MAE
• RMSE
• perceptual similarity

Roughness:

• MAE
• RMSE

Specular properties:

• MAE
• RMSE

Surface normals:

• angular error

Relighting

Relighting evaluation includes:

• image reconstruction quality
• perceptual similarity
• material consistency
• appearance consistency under unseen illumination

Uncertainty

Uncertainty evaluation includes:

• uncertainty/error correlation
• calibration
• confidence maps
• high-error/high-uncertainty overlap

Real-World Evaluation

Synthetic data provides controlled ground truth but cannot fully represent real smartphone imagery.

The project therefore includes a real-world evaluation stage.

The target input is a short smartphone video of a physical object.

A typical capture is expected to last approximately 20–60 seconds.

The intended workflow is:

Physical Object
      |
      v
Smartphone Video
      |
      v
Frame Extraction
      |
      v
Object Segmentation
      |
      v
Depth Estimation
      |
      v
Camera Pose Estimation
      |
      v
3D Gaussian Reconstruction
      |
      v
Material Decomposition
      |
      v
Uncertainty Estimation
      |
      v
Relighting

Important capture conditions include:

• sufficient object visibility
• sufficient viewpoint coverage
• reasonable camera motion
• limited motion blur
• reasonable illumination
• minimal severe occlusion

Synthetic-to-Real Transfer

Synthetic and real data serve different roles.

Synthetic data provides:

• known geometry
• known material properties
• known lighting
• known camera parameters
• pixel-aligned supervision

Real data introduces:

• sensor noise
• exposure variation
• white-balance variation
• motion blur
• compression
• lens characteristics
• uncontrolled illumination
• imperfect segmentation
• imperfect depth

The project therefore treats the synthetic-to-real gap as an explicit research problem.

Performance on synthetic data will not automatically be interpreted as equivalent real-world performance.

Limitations

Material Ambiguity

Material and illumination are strongly coupled in RGB observations.

Reflective Surfaces

Highly reflective surfaces can reproduce environmental appearance, making intrinsic material estimation difficult.

Transparent Objects

Transparent and translucent objects remain challenging because their appearance depends on geometry and surrounding illumination.

Unseen Surfaces

Surfaces that are never observed cannot be reliably reconstructed.

Depth Errors

Depth estimation errors can propagate into reconstruction and material prediction.

Camera Pose Errors

Incorrect camera poses can reduce multi-view consistency.

Synthetic-to-Real Gap

Synthetic rendering cannot perfectly reproduce real smartphone capture conditions.

Dataset Bias

ABO is a product-oriented 3D object dataset. A selected subset may therefore contain biases related to the source collection and object categories.

Data Governance and Provenance

The primary 3D object source is the Amazon Berkeley Objects dataset.

Selected 3D assets are accessed through Amazon S3 for research data preparation.

This repository does not claim ownership of the original ABO assets.

Any use or redistribution of ABO-derived data must comply with the original dataset's licensing and usage requirements.

External datasets, pretrained models and published methods used by the research remain subject to their respective licenses and citation requirements.

Users should consult the original source documentation before redistributing derived assets.

Reproducibility

The project aims to preserve sufficient information to reproduce data preparation and model experiments.

The data pipeline should maintain:

• source metadata
• object identifiers
• processing batches
• downloaded asset records
• validation results
• dataset manifests

Synthetic samples should record, where applicable:

• object identifier
• camera parameters
• lighting parameters
• material parameters
• renderer configuration
• random seed
• generated channels

Model experiments should record:

• model configuration
• dataset split
• random seed
• optimizer
• learning rate
• batch size
• training duration
• checkpoint
• evaluation metrics

Current Status

The current stage of the project focuses on establishing the 3D data infrastructure using Amazon Berkeley Objects and Amazon S3.

Current work includes:

• metadata discovery
• object selection
• candidate filtering
• S3 retrieval
• GLB validation
• asset organization
• manifest generation

The research system is being developed incrementally.

Planned development stages include:

1. synthetic multi-view rendering
2. segmentation
3. depth estimation
4. camera estimation
5. 3D Gaussian reconstruction
6. material decomposition
7. uncertainty estimation
8. material-aware Gaussian representations
9. differentiable rendering
10. relighting
11. real-world smartphone evaluation
12. quantitative and qualitative evaluation