3D Point Clouds & LiDAR Perception
Benchmarking and Adapting Indoor-Trained 3D Perception and Generation Models to Outdoor and LiDAR Point Cloud Scenes
Generated automatically from the limitations stated in 20 papers (CVPR, NeurIPS, ICML), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Nearly every 3D foundation-ish model — for detection, segmentation, registration, scene generation, and novel view synthesis — is trained and validated exclusively on indoor RGB-D or synthetic benchmarks (ScanNet, SUN RGB-D, 3D-FRONT, Habitat simulators), with explicit author acknowledgments that outdoor and LiDAR data is untested. This means the field's core 3D representations are calibrated to room-scale geometry (bounded depth, dense views, flat priors like indoor depth predictors and category vocabularies), and nobody can currently deploy them on autonomous-driving or aerial scans without unquantified degradation. The consequence is a silent incompatibility: outdoor researchers must either rebuild models from scratch or trust unmeasured transfer.
Why it matters
Quantified, reusable evidence of what breaks and why when indoor-native 3D models meet outdoor scans — plus adapted checkpoints and a benchmark suite that future model papers must report on, displacing the ScanNet-only evaluation norm.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Build an "indoor-to-outdoor transfer suite": take 5–8 representative released models spanning detection, segmentation, and reconstruction, and evaluate them, off the shelf, on outdoor benchmarks (SemanticKITTI, nuScenes, ONCE, Waymo) under matched class mappings. Measure class-wise mIoU/AP and geometry-driven failure modes (density, extent, sensor noise) versus indoor baselines, isolating which failure comes from domain shift versus sensor modality.
- 2
Interpolate a mid-domain probe: collect or reconstruct paired indoor-like/outdoor-like scans (e.g., campus-scale multi-view LiDAR, docked indoor-outdoor sequences like Aria/Tartan datasets) and measure performance as a function of scene extent, point density, and view coverage — producing transfer curves rather than a binary indoor/outdoor result.
- 3
Domain adaptation minimalism: apply test-time adaptation (e.g., entropy minimization, pseudo-label self-training) to the worst-performing models on outdoor scans and measure how much of the transfer gap can be closed without outdoor training data; report whether priors learned indoors (depth scales, room-size layouts) are repairable at inference time or baked in.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
The degradation turns out to be trivially explained by point-density and sensor-type mismatch alone (so the interesting "indoor prior" story dissolves into standard domain adaptation), or existing robust-model papers have already quietly included outdoor splits that were simply never cited in the limitation statements.
Sub-problems
Cross-Domain Robustness and Out-of-Domain Generalization in Feed-Forward 3D Gaussian Scene Priors
Current feed-forward 3D Gaussian Splatting and novel view synthesis pipelines rely on priors and models trained exclusively on curated indoor datasets such as 3D-FRONT, ScanNet++, and HyperSim. Because these systems have not been evaluated outside room-scale synthetic or scanned indoor environments, their actual degradation on outdoor, unbounded, or unstructured real-world scenes remains unquantified. Consequently, researchers cannot determine whether architectural limitations, depth-prior biases, or dataset artifacts are the primary bottleneck when deploying these models beyond indoor domains. A rigorous multi-domain robustness evaluation is necessary to establish the operational boundaries and transfer limits of existing indoor-trained 3D scene synthesis methods.
Evidence
Each paper's own statement of the limitation, verbatim.
- SpaCeFormer: Fast Proposal-Free Open-Vocabulary 3D Instance SegmentationICML 2026
Evaluated only on indoor scenes; outdoor/in-the-wild settings untested
- Scenes as Tokens: Multi-Scale Normal Distributions Transform Tokenizer for General 3D Vision-Language UnderstandingCVPR 2026
Evaluated only on ScanNet indoor scenes; no evidence on outdoor or autonomous-driving-scale 3D data despite those being cited motivations
- Flash-Mono: Feed-Forward Accelerated Gaussian Splatting Monocular SLAMICLR 2026
Performance relies on heavy pretraining with ground-truth depth and camera poses on indoor datasets (DL3DV, ScanNet++).
- EPS3D: End-to-End Feed-Forward 3D Panoptic SegmentationICML 2026
Evaluated primarily on indoor environments (ScanNet, ScanNet++, Replica), leaving open-world outdoor scene scalability unverified.
- CompetitorFormer: Mitigating Query Conflicts for 3D Instance Segmentation via Competitive StrategyCVPR 2026
Evaluated primarily on static indoor point cloud segmentation benchmarks, leaving outdoor and dynamic scenes unverified.
- Supercharging Floorplan Localization with Semantic RaysICCV 2025
Evaluated only on synthetic (Structured3D) and unfurnished panoramic-derived (ZInD) data; semantic categories limited to wall, window, door; single CPU timing may not reflect typical deployment
- LT3SD: Latent Trees for 3D Scene DiffusionCVPR 2025
Trained and evaluated only on indoor house scenes from 3D-FRONT (6,479 houses); outdoor, unstructured, or real-scan scenes are not demonstrated
- DyPolySeg: Taylor Series-Inspired Dynamic Polynomial Fitting Network for Few-shot Point Cloud Semantic SegmentationICML 2025
Evaluation limited to 100 test episodes and indoor benchmark datasets (S3DIS, ScanNet); no outdoor or larger-way settings reported
Show all 20 papers
- DepR: Depth Guided Single-view Scene Reconstruction with Instance-level DiffusionICCV 2025
Trained only on 3D-FRONT indoor furniture scenes and excludes background elements such as walls
- Cross-Modal and Uncertainty-Aware Agglomeration for Open-Vocabulary 3D Scene UnderstandingCVPR 2025
Evaluated only on indoor RGB-D benchmarks (ScanNetV2, Matterport3D) with three specific teachers; requires posed multi-view images aligned to the point cloud, so it does not apply to raw unposed point clouds or outdoor driving scenes
- AirRoom: Objects Matter in Room ReidentificationCVPR 2025
Experiments are on simulated indoor datasets built with the Habitat Simulator (Matterport3D, HM3D, Gibson, Replica) rather than real-world captures, and each room has only one reference image
- SceneCraft: Layout-Guided 3D Scene GenerationNeurIPS 2024
Requires users to supply a hand-built bounding-box layout and camera trajectory; quality depends on ScanNet++/Hypersim-style indoor training data, limiting generalization to other scene domains
- Part2Object: Hierarchical Unsupervised 3D Instance SegmentationECCV 2024
Evaluated only on indoor datasets; effectiveness on outdoor or sparse single-scan scenes without adjacent RGB frames is untested
- ControlRoom3D: Room Generation using Semantic Proxy RoomsCVPR 2024
Restricted to room-scale indoor scenes; relies on indoor-focused priors (HyperSim adapters, Matterport3D-trained CAA, ZoeDepth) and would not transfer to open outdoor scenes
- 3DET-Mamba: Causal Sequence Modelling for End-to-End 3D Object DetectionNeurIPS 2024
Only evaluated on indoor single-scan benchmarks (ScanNet, SUN RGB-D); not tested on meshes or outdoor/other 3D data types, as acknowledged by the authors.
- MM-3DScene: 3D Scene Understanding by Customizing Masked Modeling With Informative-Preserved Reconstruction and Self-Distilled ConsistencyCVPR 2023
Evaluated only on indoor scenes (ScanNetv2, SUN RGB-D, S3DIS); outdoor/LiDAR settings are not tested
- Inner-Outer Aware Reconstruction Model for Monocular 3D Scene ReconstructionNeurIPS 2023
Trained and evaluated only on indoor RGB-D datasets (ScanNet, ICL-NUIM, TUM-RGBD); no outdoor or unposed-RGB-only setting
- Geometry and Uncertainty-Aware 3D Point Cloud Class-Incremental Semantic SegmentationCVPR 2023
Evaluated only on indoor scenes (S3DIS, ScanNet); outdoor settings are left untested
- Bridging the Domain Gap: Self-Supervised 3D Scene Understanding with Foundation ModelsNeurIPS 2023
Evaluated only on indoor scene datasets (ScanNet, SUN RGB-D, S3DIS); outdoor and open-vocabulary settings are untested
- AShapeFormer: Semantics-Guided Object-Level Active Shape Encoding for 3D Object Detection via TransformersCVPR 2023
Evaluated only on indoor single-view/scene datasets (SUN RGB-D, ScanNetV2); outdoor or LiDAR detection is untested
Nearest existing work
- LiT: Unifying LiDAR "Languages" with LiDAR TranslatorNeurIPS 2024
- Instant Domain Augmentation for LiDAR Semantic SegmentationCVPR 2023
- LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor ScenesCVPR 2024
- Complete & Label: A Domain Adaptation Approach to Semantic Segmentation of LiDAR Point CloudsCVPR 2021
- An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen DomainsCVPR 2024
- 3D Semantic Segmentation in the Wild: Learning Generalized Models for Adverse-Condition Point CloudsCVPR 2023
- Single Domain Generalization for LiDAR Semantic SegmentationCVPR 2023
- Domain Generalization of 3D Semantic Segmentation in Autonomous DrivingICCV 2023
- SIRA-PCR: Sim-to-Real Adaptation for 3D Point Cloud RegistrationICCV 2023
- Scalability in Perception for Autonomous Driving: Waymo Open DatasetCVPR 2020
- Towards Explicit Geometry-Reflectance Collaboration for Generalized LiDAR Segmentation in Adverse WeatherCVPR 2025
- Walking Your LiDOG: A Journey Through Multiple Domains for LiDAR Semantic SegmentationICCV 2023
- SPG: Unsupervised Domain Adaptation for 3D Object Detection via Semantic Point GenerationICCV 2021
- BeyondMix: Leveraging Structural Priors and Long-Range Dependencies for Domain-Invariant LiDAR SegmentationNeurIPS 2025
- Robo3D: Towards Robust and Reliable 3D Perception against CorruptionsICCV 2023
Related open problems
3D Point Clouds & LiDAR Perception
An Outdoor-Scale Benchmark for 3D Scene Understanding Methods Currently Validated Only on Indoor ScanNet-Derived Data
Nearly the entire 3D perception, scene reconstruction, grounding, and question-answering stack is trained and evaluated on ScanNet-family indoor scans, so claims of scene understanding hold only within bounded, static, room-scale environments with known camera poses. As a result, no one can tell whether reported architectural advances—reasoning modules, open-vocabulary grounding, generalization arguments—transfer to outdoor, unbounded, or unsegmented scenes at all, or whether they simply exploit ScanNet's statistics. Progress reports that look cumulative within this benchmark family may be incommensurable with outdoor work that cannot reuse the same evaluation. Any paper today claiming 3D scene understanding is doing so on a single environmental regime.
3D Point Clouds & LiDAR Perception
Robust Multimodal 3D Perception Under Sensor Extrinsic and Intrinsic Miscalibration
State-of-the-art multimodal 3D perception architectures strictly require accurate extrinsic and intrinsic calibration between LiDAR and camera frames to construct point-to-pixel projections, frustum cost volumes, and cross-attention neighbourhoods. In real-world deployments, mechanical vibration, thermal expansion, and mounting shifts cause calibration parameters to drift, leading to severe geometric misalignment. Because current multimodal models hardcode these spatial transformations into their feature lifting layers, minor calibration errors cause catastrophic performance degradation or complete pipeline failure.
3D Point Clouds & LiDAR Perception
Benchmark and Training Protocols for 3D Perception from Unpaired or Calibration-Free 2D-3D Data
Nearly every current 2D-3D fusion method assumes registered multi-view imagery, known intrinsics/extrinsics, and per-frame paired data — assumptions that hold only in curated lab captures. The consequence is that point clouds collected standalone (LiDAR scans, crowdsourced reconstructions, legacy datasets) cannot benefit from image features at all, and methods cannot be deployed where a fixed calibrated rig with synchronized cameras is impractical. Progress is blocked because no established protocol or dataset exists for learning 2D-3D correspondences from unpaired, uncalibrated, or single-view inputs.