3D Point Clouds & LiDAR Perception
Benchmark and Training Protocols for 3D Perception from Unpaired or Calibration-Free 2D-3D Data
Generated automatically from the limitations stated in 11 papers (CVPR, ICCV, ECCV), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
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.
Why it matters
Methods that work on raw point clouds and arbitrary image configurations, enabling deployment on LiDAR vehicles, monocular robots, and uncalibrated consumer captures without re-engineering per rig.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Build an evaluation suite that degrades pairing: start from paired datasets (ScanNet-style), then systematically remove calibration, drop views to single-image, and jitter extrinsics; measure AP/recall curves as a function of pairing quality to quantify how much of current methods' performance is attributable to registration rather than to learned features.
- 2
Develop and test calibration-free distillation: train a point-cloud-only student network to match the features of a paired-image teacher, using self-supervised alignment losses; measure how close student performance on raw point clouds gets to the teacher's paired-input performance on standard benchmarks.
- 3
Explore test-time calibration estimation — predicting intrinsics/extrinsics from the data itself as a preprocessing module — and measure end-task recovery when its estimates substitute for ground-truth calibration.
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
Distilled point-cloud students may plateau far below paired-input performance because image features genuinely encode geometry absent from sparse point clouds, meaning the barrier is information-theoretic rather than methodological — in which case the degradation benchmark is the only durable contribution.
Sub-problems
Hand-Object Interaction and Affordance Learning Without Calibrated Multi-View Camera Rigs
Current 3D hand-object interaction and affordance learning methods require synchronized, calibrated multi-view camera arrays or fixed multi-camera rigs during training and inference for geometric triangulation. This dependence restricts research to lab capture domes and prevents these models from operating on standard egocentric head-mounted displays or monocular consumer video. Consequently, models cannot generalize to unconstrained in-the-wild environments where synchronous multi-view calibration data does not exist.
Evidence
Each paper's own statement of the limitation, verbatim.
- UST-Hand: An Uncertainty-aware Spatiotemporal Point Cloud Interaction Network for 3D Self-supervised Hand Pose EstimationCVPR 2026
Requires synchronized multi-view video with calibrated camera parameters during training to perform geometric triangulation and projection.
- Selective Contrastive Learning for Weakly Supervised Affordance GroundingICCV 2025
Requires paired multi-view data (E=3 exocentric images per egocentric image), which may not exist in real deployment settings.
- Point Cloud Self-supervised Learning via 3D to Multi-view Masked LearnerICCV 2025
Requires camera pose information for the 3D-to-2D feature projection, which may not be available for arbitrary point cloud data
- IAAO: Interactive Affordance Learning for Articulated Objects in 3D EnvironmentsCVPR 2025
Requires posed multi-view images of the scene at two distinct joint states captured in advance, limiting purely passive or single-state applicability
- Weakly Supervised Point Cloud Semantic Segmentation via Artificial OracleCVPR 2024
Requires registered multi-view images, depth maps, and camera projection matrices during training, so it does not apply to point clouds without paired imagery
- Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose EstimationCVPR 2024
Requires a sufficient number of unlabeled dual-view image pairs (performance converges only at N ≥ 1000), so it needs paired synchronous dual-camera data from a fixed rig
- Open3DIS: Open-Vocabulary 3D Instance Segmentation with 2D Mask GuidanceCVPR 2024
Requires RGB-D sequences with known camera intrinsics/extrinsics per frame; the method cannot run from point clouds alone (2D-only mode drops AP head/com recall relative to combined)
- Beyond Viewpoint: Robust 3D Object Recognition under Arbitrary Views through Joint Multi-Part RepresentationECCV 2024
Requires multi-view rendered or real multi-view image inputs; no evaluation on raw point clouds or single-view inputs
Show all 11 papers
- PointDC: Unsupervised Semantic Segmentation of 3D Point Clouds via Cross-Modal Distillation and Super-Voxel ClusteringICCV 2023
Requires multi-view RGB images with known camera intrinsic and extrinsic parameters paired to the point clouds (ScanNet-v2 2D data, 2D-3D-S), so it does not apply to pure geometry-only point clouds or unknown-calibration settings
- PiMAE: Point Cloud and Image Interactive Masked Autoencoders for 3D Object DetectionCVPR 2023
Requires calibrated camera intrinsics/extrinsics and paired, aligned RGB-D data (SUN RGB-D style), so it does not apply to unpaired or uncalibrated 2D/3D data
- POEM: Reconstructing Hand in a Point Embedded Multi-View StereoCVPR 2023
Depends on multi-view calibrated camera setups (known intrinsics/extrinsics); not applicable to monocular input
Nearest existing work
- LiREC-Net: A Target-Free and Learning-Based Network for LiDAR, RGB, and Event CalibrationCVPR 2026
- BEVFusion: A Simple and Robust LiDAR-Camera Fusion FrameworkNeurIPS 2022
- DeepFusion: Lidar-Camera Deep Fusion for Multi-Modal 3D Object DetectionCVPR 2022
- UniCal: Unified Neural Sensor CalibrationECCV 2024
- PointMBF: A Multi-scale Bidirectional Fusion Network for Unsupervised RGB-D Point Cloud RegistrationICCV 2023
- Boosting 3D Object Detection by Simulating Multimodality on Point CloudsCVPR 2022
- Diff2I2P: Differentiable Image-to-Point Cloud Registration with Diffusion PriorICCV 2025
- Leveraging Vision-Centric Multi-Modal Expertise for 3D Object DetectionNeurIPS 2023
- Noise-Aware Unsupervised Deep Lidar-Stereo FusionCVPR 2019
- Differentiable Registration of Images and LiDAR Point Clouds with VoxelPoint-to-Pixel MatchingNeurIPS 2023
- TransFusion: Robust LiDAR-Camera Fusion for 3D Object Detection With TransformersCVPR 2022
- PointAugmenting: Cross-Modal Augmentation for 3D Object DetectionCVPR 2021
- GraphI2P: Image-to-Point Cloud Registration with Exploring Pattern of Correspondence via Graph LearningCVPR 2025
- DeepI2P: Image-to-Point Cloud Registration via Deep ClassificationCVPR 2021
- FULLER: Unified Multi-modality Multi-task 3D Perception via Multi-level Gradient CalibrationICCV 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
Benchmarking and Adapting Indoor-Trained 3D Perception and Generation Models to Outdoor and LiDAR Point Cloud Scenes
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.
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.