Source-linked AI summary
DIODE: A Dense Indoor and Outdoor DEpth Dataset
Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z. Dai, Andrea F. Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R. Walter, Gregory Shakhnarovich
TL;DR
Existing RGBD datasets lack extensive, diverse ground-truth depth collected with the same sensor across indoor and outdoor settings. DIODE introduces a large dataset using a survey-grade scanner and unified imaging setup, providing dense, accurate, long-range measurements and enabling unified depth evaluation. Its scope remains bounded by LiDAR-specific artifacts and a deprecated crop version.
Problem
Public datasets lack extensive, diverse depth ground truth measured with the same sensor in both indoor and outdoor settings.
Method
DIODE collects diverse indoor and outdoor RGBD scenes with a FARO Focus S350 survey-grade laser scanner whose color camera is aligned with depth returns.
Results
DIODE provides larger scene variety, higher image and depth resolution, and denser, more accurate measurements within a unified indoor–outdoor framework.
Takeaways & Limitations
DIODE is intended to support more realistic evaluation of depth prediction methods and facilitate progress toward general depth estimation methods.
Takeaways & Limitations
LiDAR measurements can produce erroneous returns on specular objects, see through glass, and create RGB–depth inconsistencies in darkness.
Abstract
from arXiv · showhide
We introduce DIODE, a dataset that contains thousands of diverse high resolution color images with accurate, dense, long-range depth measurements. DIODE (Dense Indoor/Outdoor DEpth) is the first public dataset to include RGBD images of indoor and outdoor scenes obtained with one sensor suite. This is in contrast to existing datasets that focus on just one domain/scene type and employ different sensors, making generalization across domains difficult. The dataset is available for download at http://diode-dataset.org
1. Introduction
Existing RGBD datasets are limited by noisy or sparse sensing, domain-specific collection, and a lack of diverse indoor–outdoor coverage. DIODE addresses these gaps with dense, accurate measurements from one sensing and imaging setup across both scene types.
- RGBD datasets rely on noisy, sparse, expensive range sensors because depth cannot be crowd-sourced like image labels.
- Available RGBD datasets primarily cover either indoor or outdoor scenes because their sensors and collection settings are domain-specific.
- The lack of large, diverse real-world RGBD datasets comparable to ImageNet and COCO has hindered progress in 2.5D and 3D vision.
- Depth supports robotics tasks such as mapping, localization, and obstacle avoidance, as well as augmented and virtual reality.
- DIODE uses a survey-grade laser scanner to provide dense, high-resolution depth and accurate surface normals for unified indoor and outdoor scenes.
2. Related Work
Prior RGBD datasets differ in scene coverage, sensing quality, scale, and intended use, often focusing on one domain or task. DIODE combines broad scene variety, high-resolution accurate measurements, and unified indoor–outdoor coverage for depth-related learning and evaluation.
- Most RGBD datasets consist exclusively of indoor or outdoor scenes and are often tied to applications such as residential modeling or autonomous driving.
- Outdoor scenes: KITTI provides sparse LiDAR depth for street scenes around Karlsruhe, while its self-driving focus yields lower outdoor scene diversity than DIODE.
- Outdoor scenes: Cityscapes has no depth ground truth, while Synthia and Sintel use synthetic depth and Megadepth lacks ground-truth depth and scale.
- Outdoor scenes: Make3D offers diverse outdoor scenes but has very low-resolution depth maps compared with DIODE.
- Outdoor scenes: ETH3D includes indoor and outdoor scenes with similar sensing diversity, but is intended for benchmarking and is an order of magnitude smaller than DIODE.
- Outdoor scenes: 3D Movies provides many diverse frames and dynamic objects, but its depth is approximate and lacks scale.
- Indoor scenes: NYUv2 supplies indoor RGBD data with sparse, noisy Kinect returns that are generally inpainted and smoothed before monocular depth estimation.
- Indoor scenes: Matterport3D and ScanNet provide dense indoor depth images, but their SLAM-rendered maps are noisier and lower-resolution than DIODE.
3. The DIODE Dataset
DIODE is designed around diverse indoor and outdoor scenes, dense accurate long-range depth, and a survey-grade scanner that supports aligned RGB and depth capture. Its processing pipeline converts scans into rectified crops, depth maps, normals, and curated validity masks while preserving broad depth and scene diversity.
- Design desiderata: DIODE targets diverse indoor and outdoor scenes, dense short-, mid-, and long-range measurements, and highly accurate depth.Indoor examples include homes, offices, lecture halls, and communal spaces; outdoor examples include streets, parks, forests, and river banks.
- Data acquisition: The FARO Focus S350 provides 0.6 m–350 m depth coverage, error as low as 1 mm, 0.009° angular resolution, and coaxially aligned RGB panoramas.The scanner is equally suited to indoor and outdoor environments and produces dense point clouds with minimal RGB-depth baseline.
- Scanning settings: Scans use 360° horizontal and 150° vertical fields of view with HDR RGB, producing approximately 11-minute scans and high-resolution panoramas with associated colored 3D points.Each intermediate panorama is approximately 20700 × 8534, although specular surfaces, sky, and regions seen through glass or darkness may lack consistent depth or appearance.
- Scene diversity: DIODE scenes vary in both location and composition, including nearby objects outdoors and distant objects indoors beyond the range typical of short-range indoor sensors.The dataset includes small offices, residential buildings, hiking trails, meeting halls, parks, streets, and parking lots across three cities.
- Image and depth extraction: The pipeline extracts 768 × 1024 rectified crops from 216 viewing directions and projects RGB and depth by matching rays to panorama pixels and point-cloud returns.Depth uses returns within 0.5° and a robust mean over the median 80% of matched depths; empty matches are coded as depth 0.
- Curation and normals: Normals are estimated with RANSAC planes over local point neighborhoods, then oriented toward the scanner and rectified for each crop.The dataset is manually curated to remove problematic dynamic, empty, and featureless crops, while automated validity masks reject spurious depth returns.
- Dataset statistics: DIODE provides returns for most color pixels, fine depth edges, thin objects, and depth values beyond 100 m, with outdoor distributions more diffuse than KITTI’s.Only the smaller, lower-resolution Make3D dataset approaches DIODE’s depth-value diversity.
4. Experiments
The experiments establish DenseDepth as a baseline for monocular depth estimation on DIODE and compare training across indoor, outdoor, and combined subsets. Combined training performs best overall, although domain-specific validation reveals different trade-offs.
- Baseline setup: DenseDepth is evaluated on DIODE/Indoor, DIODE/Outdoor, and DIODE/All using full-resolution RGB inputs and half-resolution predicted depth.All models are trained for 30 epochs with Adam and evaluated after 2× upsampling to ground-truth resolution.
- Results: DIODE/All outperforms DIODE/Indoor on indoor validation.The paper attributes this partly to the outdoor subset having roughly twice as many images and containing objects across a wide range of distances.
- Results: Performance slightly degrades on outdoor validation when training uses DIODE/All rather than the outdoor subset.The paper suggests indoor scenes mostly contain objects within approximately 50 m of the camera, which may limit outdoor-range specialization.
- Results: DIODE/All performs better on the entire validation set than models trained on either indoor or outdoor subsets.This result supports using the combined dataset for evaluation across both scene domains.
5. Conclusion
DIODE’s dense, accurate depth data and unified indoor-outdoor collection framework are intended to support more realistic evaluation and progress toward general depth estimation methods.
- Conclusion: DIODE’s dense and accurate depth measurements, together with its unified indoor-outdoor framework, are expected to enable more realistic evaluation of depth prediction methods.The authors also plan to expand the dataset with more locations and greater weather and seasonal variety.