DF3DV-1K dataset ships 1,048 scenes for distractor-free 3D reconstruction
DF3DV-1K is a large-scale dataset from researchers at multiple institutions that pairs clean and cluttered image sets across 1,048 real-world scenes to benchmark radiance field methods that remove distractors.
DF3DV-1K is a large-scale real-world dataset from researchers Cheng-You Lu, Yi-Shan Hung, Wei-Ling Chi, Hao-Ping Wang, Charlie Li-Ting Tsai, and Yu-Cheng Chang that pairs clean and cluttered image sets across 1,048 scenes to benchmark distractor-free radiance field methods. The dataset contains 89,924 images captured with consumer cameras, spanning 128 distractor types and 161 scene themes across indoor and outdoor environments. Each scene includes both clean reference images and cluttered versions with distractors — objects, people, or visual noise that complicate novel view synthesis.
Radiance fields have enabled photorealistic 3D reconstruction from 2D images, but most large-scale datasets focus on scene-specific reconstruction without addressing distractors. DF3DV-1K fills that gap by providing paired data that lets researchers measure how well methods handle real-world clutter. The authors also curated DF3DV-41, a 41-scene subset designed to stress-test methods under challenging scenarios.
Benchmark results
The team benchmarked nine recent distractor-free radiance field methods and 3D Gaussian Splatting on DF3DV-1K, identifying which methods handle clutter most robustly and which scenarios remain hardest to solve. Beyond benchmarking, they fine-tuned a diffusion-based 2D enhancer on the dataset and achieved average improvements of 0.96 dB PSNR and 0.057 LPIPS on the held-out DF3DV-41 set and the On-the-go dataset. The dataset and leaderboard are available at the project page.
