DF3DV-1K: A Large-Scale Dataset and Benchmark for Distractor-Free Novel View Synthesis

2026-04-16 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

DF3DV-1K is a new, large-scale, real-world dataset and benchmark designed to advance distractor-free novel view synthesis. It comprises 1,048 indoor and outdoor scenes, totaling 89,924 images captured with consumer cameras, featuring both clean and cluttered image sets per scene. The dataset encompasses 128 distractor types and 161 scene themes. A curated subset, DF3DV-41, includes 41 scenes specifically designed to test method robustness under challenging conditions like semantically similar or fluid distractors and nighttime scenes. The benchmark evaluates nine recent distractor-free radiance field methods and 3D Gaussian Splatting, identifying AsymGS and RobustSplat as the most robust. Additionally, DF3DV-1K was used to fine-tune DI2FIX, a diffusion-based 2D enhancer, which improved radiance field methods by 0.96 dB PSNR and reduced LPIPS by 0.057 on held-out and On-the-go datasets.

Key takeaway

For AI Scientists and Computer Vision Engineers developing or evaluating distractor-free novel view synthesis methods, DF3DV-1K offers a critical resource. You should leverage this large-scale, diverse dataset and its challenging subsets (DF3DV-41) to rigorously benchmark your models, particularly against difficult scenarios like fluid distractors or low-light conditions. Consider integrating enhancers like DI2FIX, fine-tuned on DF3DV-1K, to improve rendering quality and push beyond scene-specific solutions.

Key insights

DF3DV-1K provides a large-scale, diverse dataset and benchmark for distractor-free novel view synthesis.

Principles

• Dataset scale and diversity improve model robustness.
• Systematic scenario design reveals method limitations.
• Moderate data degradation levels optimize training.

Method

DF3DV-1K data acquisition involves predefined scene themes, distractor types, and viewpoint coverage, using consumer cameras with automatic settings. Curation includes manual review, COLMAP for pose estimation, and instant-ngp for quality verification.

In practice

• AsymGS and RobustSplat are top-performing methods.
• Semantically similar and fluid distractors are challenging.
• DI2FIX enhances radiance field rendering quality.

Topics

• DF3DV-1K Dataset
• Distractor-Free Novel View Synthesis
• Radiance Fields Benchmarking
• 3D Gaussian Splatting
• DI2FIX 2D Enhancer

Best for: AI Scientist, Computer Vision Engineer, Research Scientist

Related on AIssential

• DF3DV-1K: A Large-Scale Dataset and Benchmark for Distractor-Free Novel View Synthesis — DF3DV-1K is a new dataset for benchmarking distractor-free novel view synthesis methods.
• DVSM: Decoder-only View Synthesis Model Done Right — Decoder-only view synthesis models, like DVSM, can outperform encoder-decoder architectures with fewer parameters and aligned features.
• EditSSC: Toward Editable Semantic Occupancy Scenes with Unconditional Diffusion Models — EditSSC repurposes 2D latent diffusion for 3D semantic scene generation, enabling training-free editing capabilities.
• VisDom: Sparse Novel View Synthesis with Visible Domain Constraint — VisDom uses a multi-view visibility constraint to improve sparse novel view synthesis by reducing overfitting and artifacts.
• DUALVISION: RGB-Infrared Multimodal Large Language Models for Robust Visual Reasoning — Integrating infrared and RGB data via localized cross-attention significantly boosts MLLM robustness in degraded visual conditions.
• QueryGaussian: Scalable and Training-Free Open-Vocabulary 3D Instance Retrieval — QueryGaussian enables scalable, training-free 3D instance retrieval by decoupling semantic understanding from geometric representation, overcoming OOM issues.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.