GhostObjects in Sensing, AR & Physics
- GhostObjects are entities defined by discrepancies between perceived signals and physical reality, appearing in automotive sensing, AR robotics, optical imaging, and theoretical physics.
- In automotive sensing, ghost objects are false detections from multi-path reflections in radar/LiDAR systems, with studies showing significant impacts on false positive rates and detection accuracy.
- In augmented reality robotics, GhostObjects are world-aligned virtual twins used to instruct robots through precise, life-size overlays that replicate real-world objects.
Searching arXiv for papers on “GhostObjects” and related uses to ground the article. arxiv_search(query="GhostObjects OR \"ghost objects\" OR GhostObjects", max_results=10, sort_by="relevance") Refining the search to capture the robotics usage and sensing-related “ghost objects” literature. arxiv_search(query="all:\"ghost objects\" automotive radar lidar augmented reality robotics", max_results=10, sort_by="submittedDate") GhostObjects is a domain-specific term whose meaning varies markedly across contemporary research. In automotive sensing, ghost objects are false detections caused by multi-path reflections in radar or LiDAR; in augmented reality robotics, GhostObjects are world-aligned, life-size virtual twins of physical objects used for robot instruction; and in adjacent literatures, closely related “ghost” terminology refers to optical ghosting artifacts, hallucinated objects in multimodal models, and exotic gravitational configurations sustained by negative energy or phantom fields (Kraus et al., 2024, Wang et al., 14 Aug 2025, Tanoglidis et al., 2021, Parast et al., 29 Sep 2025, Guilabert et al., 26 Dec 2025).
1. Semantic scope across research domains
The literature does not use GhostObjects as a single unified technical term. Instead, it names several families of entities that are operationally present to a sensing, control, or theoretical framework while being absent, indirect, virtual, or nonclassical.
| Domain | Meaning | Representative work |
|---|---|---|
| Automotive radar/LiDAR | False detections from multi-path reflections | (Kraus et al., 2024, Ikeda et al., 30 Mar 2026) |
| Augmented reality robotics | World-aligned, life-size virtual twins of physical objects | (Wang et al., 14 Aug 2025) |
| Optical sensing | Ghosting artifacts, scattered-light artifacts, lens-flare ghosts | (Tanoglidis et al., 2021, Man et al., 2020) |
| Theoretical physics | Ghost stars, ghost wormholes, phantom-field configurations | (Guilabert et al., 26 Dec 2025, Bolokhov et al., 2015) |
This suggests a common conceptual thread: a “ghost object” is typically an object-like entity that materially affects inference, action, or geometry without corresponding to an ordinary physical object in the relevant sense. In sensing, the discrepancy is between perception and scene geometry; in AR robotics, between physical objects and controllable virtual twins; in gravitation, between ordinary matter models and geometries sustained by negative energy density or phantom kinetic terms.
2. Automotive sensing: multi-path ghosts, peak-level annotations, and false-positive control
In automotive radar, ghost objects are false detections produced by multi-path reflections rather than direct line-of-sight returns. The underlying cause is tied to the large wavelength of automotive radar, approximately $4$ mm at $77$ GHz, which makes many environmental surfaces such as roads, walls, guard rails, and parked cars act as mirrors. The result is that radar may report points or clusters at locations where no real object exists, creating a major source of false positive detections in the perception pipeline (Kraus et al., 2024).
The 2024 Radar Ghost Dataset formalizes this problem with detailed manual semantic instance annotations. It distinguishes two fundamental reflection types, following Liu2016b: Type-1, where the last bounce is on the real object, and Type-2, where the last bounce is on the reflective surface. It also categorizes paths by bounce order, including First-order (Real), Second-order, and Third-order. The label set includes Real object, MP-12, MP-22, MP-23, and OMP. The dataset contains about 1 million manual annotations across 111 sequences and 21 scenarios, with approximately 100,000 points labeled as ghost objects. The main annotated objects are vulnerable road users—pedestrians and cyclists—while cars, trucks, and other vehicles are included if present (Kraus et al., 2024).
Two baseline identification strategies are evaluated on that dataset. The first is PointNet++-based Deep Semantic Segmentation, adapted for radar point clouds and using amplitude, Doppler velocity, and a relative timestamp as per-point features, with class balancing defined by
The second is instance-level grouping, implemented either by SGPN (Similarity Group Proposal Network) or by DBSCAN-based Clustering after semantic segmentation. Evaluation uses Average Precision (AP) and mean AP (mAP) at IoU thresholds of $0.3$ and $0.5$, together with semantic F1 scores. Selected results report Real objects (Obj) F1: 95.4\%, ghost-class F1 scores in the 52–72\% range, mAP (SGPN): (IoU 0.3), 67.5\% (IoU 0.5), and mAP (DBSCAN): (0.3), 66.2\%) (0.5). A confusion analysis attributes 40–60\% of false positives in detection to ghost objects, with MP-12 causing up to 36\% of FPs in some configurations (Kraus et al., 2024).
An earlier radar study concentrated on type-2, third-bounce ghosts and used a large-scale dataset with 25 different scenarios, each repeated four times, for roughly 100 recordings and 63 minutes of data. Two $77$ GHz automotive radars were used, and labels included Pedestrian, Cyclist, Ghost Pedestrian, Ghost Cyclist, and Background. In the joint object-plus-ghost setting, the reported ghost object F1-score is approximately 55\%, while real object F1-score remains approximately 94–95\%. The paper also reports a reduction in ghost-related false positives, with one example dropping from 1268 to 658 when ghost labels are included in training (Kraus et al., 2020).
The LiDAR analogue of this problem is developed in Ghost-FWL, where ghost points are defined as false reflections caused by multi-path laser returns from glass and reflective surfaces. The paper introduces Ghost-FWL, described as the first and largest annotated mobile FWL dataset for ghost detection and removal, with 24K frames across 10 diverse scenes and 7.5 billion peak-level annotations, stated to be 100x larger than existing annotated FWL datasets. It establishes a full-waveform baseline and proposes FWL-MAE, a masked autoencoder tailored to FWL data. Reported downstream gains include 66\% trajectory error reduction for LiDAR-based SLAM and 50x false positive reduction for 3D object detection after ghost removal (Ikeda et al., 30 Mar 2026).
A separate line of work treats ghost objects as adversarially spoofed LiDAR detections. Shadow-Catcher introduces 3D shadows as a semantically meaningful physical invariant for validating objects. On KITTI, it reports more than 94\% accuracy in identifying anomalous shadows for vehicles, pedestrians, and cyclists, processing times between 0.003s-0.021s per object on commodity hardware, and a 2.17x speedup compared to prior work (Hau et al., 2020). Taken together, these studies establish ghost objects as both a natural consequence of wave propagation and a security-relevant attack surface in autonomous vehicle sensing.
3. Optical and imaging uses: ghosting artifacts, scattered light, and remote perception attacks
In astronomical imaging, “ghosts” are undesirable reflections and scattered-light artifacts that contaminate wide-field survey data. DeepGhostBusters studies this problem on Dark Energy Survey (DES) images and uses Mask R-CNN to detect and localize three artifact categories: Rays, Bright ghosts, and Faint ghosts. The dataset comprises 2,000 images manually annotated by expert astronomers, each image being pixels and grayscale. The model is an adaptation of the Matterport implementation in Keras/TensorFlow, with a ResNet-101 backbone and Feature Pyramid Network (FPN), initialized from MS COCO and fine-tuned on DES data (Tanoglidis et al., 2021).
The paper evaluates both CCD-based masking quality and object detection quality. It reports that a multi-step pipeline combining Mask R-CNN segmentation with a classical CNN classifier reduces false positives by a factor of approximately . Relative to a DES Ray-Tracing baseline, it reports a substantial increase in recall and F1 score, including Mask R-CNN F1 = 72.5\% vs Ray-Tracing F1 = 55.4\% for Rays+Bright at low area threshold. Inference time is given as approximately $77$0 seconds per image on GPU (Tanoglidis et al., 2021).
In camera security and autonomy, ghost effects can be weaponized rather than merely mitigated. GhostImage demonstrates remote perception attacks against camera-based classification systems by exploiting lens flare/ghost effects and auto-exposure control. A projector shines carefully computed light patterns toward the camera lens, creating a ghost image at a distinct image location and altering the captured scene in a way that can create or modify perceived objects. The attack is optimized with adversarial machine learning and an end-to-end projector-camera channel model. The study reports that, depending on projector-camera distance, attack success rates can reach as high as 100\% and under targeted conditions; in outdoor tests, the attack failed due to sunlight overpowering the projector during daylight (Man et al., 2020).
A broader optical context uses “ghost” in a different sense: ghost imaging. In “Metasurface-Assisted Quantum Ghost Discrimination of Polarization Objects”, specially designed metasurfaces and quantum-entangled or classically-correlated photons allow identification of polarization objects and arbitrary orientation angles using four or fewer correlation measurements (Vega et al., 2021). In “Counterfactual Ghost Imaging”, the object can be imaged such that, in the infinite limit, no photons ever go to the imaged object, with over an order of magnitude reduction in absorbed intensity relative to previous protocols (Hance et al., 2020). These are not ghost objects as false positives, but they show that optical ghost terminology spans both artifacts and deliberately engineered nonlocal measurement schemes.
4. GhostObjects in augmented reality robotics
In AR-mediated robot instruction, GhostObjects has a capitalization-specific and constructive meaning. Wang and Abtahi define GhostObjects as world-aligned, life-size virtual twins of physical objects, rendered and manipulated in augmented reality to specify robot goals and spatial parameters (Wang et al., 14 Aug 2025).
The implementation uses Meta Quest’s Space Setup to scan the environment and establish spatial anchors with object bounding boxes. The Meta Mixed Reality Utility Kit (MRUK) accesses those anchors and calculates their 6-DoF poses. When selected, GhostObjects are spawned at the same real-world poses as their physical counterparts, producing precise overlay without repeated manual calibration. They are rendered as life-size, visually matched representations, including color, shape, and state, such as the fill level of a bottle, and they interact with a physics-enabled environment (Wang et al., 14 Aug 2025).
Interaction is object-centric. Users can perform Raycast Selection by projecting a ray from a hand-held controller, or Lasso Multi-Object Selection by drawing a 3D lasso on surfaces and selecting all virtual objects whose bounding boxes intersect the lassoed volume. Manipulation includes Grab, Group Move, and Modify Properties for state-transformable objects. A Snap-to-Default Feature provides a user-configurable default pose; when a GhostObject is moved, an arched trajectory line connects current and default positions, and releasing the object along that trajectory snaps it back to the default pose (Wang et al., 14 Aug 2025).
The paper frames these affordances as a way to move beyond Programming by Demonstration and teleoperation toward direct specification of physical goals. Example scenarios include home tidying, where foam blocks are lasso-selected and snapped into a basket, and filling bottle, where the user selects a bottle’s GhostObject, chooses “Fill”, and raises the controller to increase the displayed fill level (Wang et al., 14 Aug 2025). The work is explicitly presented as a system introduction and walkthrough; it reports no formal user study or quantitative experimental results at this stage. Its significance lies in defining GhostObjects as collocated virtual proxies through which humans can instruct robots with precise spatial and state-level intent.
5. Related AI usage: hallucinated objects and model-induced visual cues
A related, though not terminologically identical, usage appears in multimodal AI research on object hallucination. GHOST (Generating Hallucinations via Optimizing Stealth Tokens) is a fully automatic method for generating visual-natural, object-free images that cause multimodal LLMs to hallucinate a specified absent object. The method optimizes in CLIP embedding space, uses a trainable mapper to bridge to the target MLLM vision-token space, and then conditions Stable Diffusion unCLIP on the optimized embedding. An OWLv2 detector filters out synthesized images that actually contain the target object (Parast et al., 29 Sep 2025).
Across evaluated models, the paper reports hallucination success rates of 29.9\% for Qwen2.5-VL, 28.1\% for LLaVA-v1.6, and 32.4\% for GLM-4.1V, compared to approximately 0.1\% in prior data-driven discovery methods. Human evaluation reports that 89\% of annotators for LLaVA and 86.3\% for Qwen2.5-VL agree that the generated images do not contain the target object. The study also reports cross-model transfer, including Qwen2.5-VL $77$1 GPT-4o: 66.5\% hallucination rate, and mitigation through fine-tuning, with POPE F1 improving from 88.5 to 90.7 and no regression in VQAv2 and COCO captioning (Parast et al., 29 Sep 2025).
This is not a GhostObjects framework in the AR sense, nor a sensor ghost in the automotive sense. A plausible implication is that modern AI research increasingly treats “ghost objects” as a perception failure mode: objects that exist for the model but not for a human observer or the underlying image. That places object hallucination alongside radar ghosts, LiDAR ghosts, and optical ghosting as a broader class of epistemic discrepancies between signal and world.
6. Gravitational ghost configurations and exotic compact objects
In general relativity and related gravitation research, “ghost” denotes matter fields or compact configurations with nonstandard energy properties rather than sensing artifacts. “Traversable ghost wormholes” studies ghost stars, defined as compact configurations with arbitrarily small total mass, and extends the idea to ghost wormholes, whose total quasilocal mass vanishes. The analysis emphasizes that negative energy density is unphysical in conventional stellar models but arises naturally in traversable wormholes, where violation of the null energy condition is required at the throat. The paper shows that when the ghost condition is extended beyond spherical symmetry and applied to the Hawking mass, topological obstructions arise. It also presents a Casimir-like traversable wormhole and analyzes its Penrose-Carter diagram (Guilabert et al., 26 Dec 2025).
A distinct but related line studies phantom scalar fields as sources of wormholes and regular black holes. “Visible, invisible and trapped ghosts as sources of wormholes and black universes” constructs globally regular static, spherically symmetric solutions involving scalar and electromagnetic fields, with ghosts, trapped ghosts, and invisible ghosts. Here, ghosts are scalar fields with negative kinetic energy; trapped ghosts have negative kinetic energy only in a strong-field region; invisible ghosts are phantom scalar fields that decay sufficiently rapidly in the weak-field region. The resulting configurations may contain from zero to four Killing horizons (Bolokhov et al., 2015).
“Magnetic wormholes and black universes with invisible ghosts” further develops the invisible-ghost construction by combining a canonical scalar field with a phantom scalar field that becomes negligible at infinity. The motivation is to confine NEC violation to the throat or bounce region while avoiding the stability problems associated with sign-changing kinetic terms in trapped-ghost models. The paper again reports configurations with different numbers of Killing horizons, from zero to four, including traversable wormholes and regular black holes, especially black universes (Bronnikov et al., 2015).
These gravitational usages are terminologically distinct from sensing and AR robotics. Nevertheless, they preserve the same structural motif: an object-like or geometry-supporting entity that exists outside the regime of ordinary physical interpretation. Across fields, GhostObjects therefore names not one object class but a family of research concepts defined by mismatch—between measurement and scene, virtual proxy and physical object, model belief and image content, or spacetime geometry and standard matter.