PaperlayerView exact revision on arXiv

Source-linked AI summary

Review on Panoramic Imaging and Its Applications in Scene Understanding

Shaohua Gao, Kailun Yang, Hao Shi, Kaiwei Wang, Jian Bai

arXiv:2205.05570v2cs.CVcs.ROeess.IVphysics.optics

TL;DR

Panoramic imaging must balance broad environmental coverage with resolution, blind-area control, compactness, and multidimensional perception. This review surveys system architectures, emerging optical technologies, and scene-understanding applications. It concludes that panoramic systems have substantial application potential, while freeform optics still face design and fabrication challenges.

  • Problem

    Panoramic imaging must meet competing demands for field of view, resolution, no blind area, multidimensional perception, and lightweight compact designs, while prior panoramic surveys do not cover the field’s recent progress.

    Source: §I · p. 1

  • Method

    The review analyzes panoramic optical systems and downstream perception applications, covering architectures, optical technologies, and scene understanding.

    Source: §I · pp. 1–2, §VI · p. 25

  • Results

    The review identifies panoramic imaging’s application potential in autonomous driving and robotics and summarizes advantages, progress, limitations, challenges, and directions for improvement.

    Source: §VI · p. 25

  • Takeaways & Limitations

    The review presents panoramic imaging as a promising approach for environmental perception, with optical and algorithmic challenges remaining.

    Source: §VI · p. 25

    Freeform optics still need faster, more accurate, and more general direct point-by-point optimization, as well as improved fabrication, testing, and assembly accuracy and efficiency.

    Source: A. Freeform Surface · p. 15

Abstract

from arXiv · show

With the rapid development of high-speed communication and artificial intelligence technologies, human perception of real-world scenes is no longer limited to the use of small Field of View (FoV) and low-dimensional scene detection devices. Panoramic imaging emerges as the next generation of innovative intelligent instruments for environmental perception and measurement. However, while satisfying the need for large-FoV photographic imaging, panoramic imaging instruments are expected to have high resolution, no blind area, miniaturization, and multidimensional intelligent perception, and can be combined with artificial intelligence methods towards the next generation of intelligent instruments, enabling deeper understanding and more holistic perception of 360-degree real-world surrounding environments. Fortunately, recent advances in freeform surfaces, thin-plate optics, and metasurfaces provide innovative approaches to address human perception of the environment, offering promising ideas beyond conventional optical imaging. In this review, we begin with introducing the basic principles of panoramic imaging systems, and then describe the architectures, features, and functions of various panoramic imaging systems. Afterwards, we discuss in detail the broad application prospects and great design potential of freeform surfaces, thin-plate optics, and metasurfaces in panoramic imaging. We then provide a detailed analysis on how these techniques can help enhance the performance of panoramic imaging systems. We further offer a detailed analysis of applications of panoramic imaging in scene understanding for autonomous driving and robotics, spanning panoramic semantic image segmentation, panoramic depth estimation, panoramic visual localization, and so on. Finally, we cast a perspective on future potential and research directions for panoramic imaging instruments.

I. INTRODUCTION

Panoramic imaging offers broad environmental perception but must balance field of view, resolution, blind-area control, compactness, and multidimensional sensing. This review surveys system principles, architectures, emerging optical technologies, and scene-understanding applications.

I. INTRODUCTION

  • Existing panoramic imaging surveys do not cover the field’s flourishing progress over the last 20 years, motivating this review’s account of panoramic imaging and its applications.The review also identifies competing requirements in field of view, resolution, blind-area control, multidimensional sensing, and lightweight design.Source: §I · p. 1
  • Freeform surfaces and metasurfaces have reshaped panoramic systems, while multidimensional imaging modalities have broadened their application roles.The review covers system concepts and challenges, architectures, imaging modalities, emerging optical technologies, and scene-understanding applications.Source: §I · pp. 1–2

II. KEY PARAMETERS OF PANORAMIC IMAGING SYSTEMS

  • Panoramic-system design weighs field of view, wavelength, F-number, focal length, resolution, imaging quality, and volume against application requirements.Potential solutions to these trade-offs can support the design and manufacture of systems with higher parameter performance.Source: §II · p. 2

A. Definition of Parameters

The review defines key panoramic-optics parameters and explains how they characterize coverage, light capture, image formation, image quality, and compactness.

A. Definition of Parameters

  • A fisheye optical system can usually reach a field of view of 180° or more, capturing more of the surrounding scene at once.The review uses full-angle field of view as the usual practical evaluation standard and distinguishes systems with a blind area.Source: A. Definition of Parameters · pp. 2–3
  • Visible light is usually defined as 400∼700 nm, while infrared and ultraviolet panoramic systems extend sensing beyond the human-visible spectrum.The review gives typical infrared wavelengths as 700 nm to 1 mm and ultraviolet wavelengths as 10–400 nm.Source: A. Definition of Parameters · p. 3
  • F-number is focal length divided by entrance-pupil diameter; smaller values admit more light but generally reduce depth of field and increase design difficulty.The review notes that small-F-number systems can suit low-light or moving-object imaging, whereas larger F-numbers provide greater depth of field with reduced illumination.Source: A. Definition of Parameters · p. 3
  • The equidistant projection model relates image height y to focal length f and half field of view θ as y = f · θ.The review describes five classical projection models and notes that the sensor must be matched to the optical system.Source: A. Definition of Parameters · p. 3
  • Resolution refers to sensor pixel-array counts, while imaging quality is evaluated using spot diagrams and modulation transfer function alongside aberration correction.The review also calls for uniform relative illumination and low distortion to meet perception-accuracy requirements.Source: A. Definition of Parameters · pp. 3–4
  • The panoramic compactness ratio compares the system’s maximum lateral diameter with the maximum diameter of its sensor imaging circle.Miniaturized and lightweight systems are suited to space- and weight-constrained scenarios.Source: A. Definition of Parameters · p. 4

B. Key Parameters Trade-offs and Potential Solutions

Large field of view complicates aberration correction, and panoramic architectures trade optical performance against size, cost, and system complexity. The review describes conventional and newer architectures alongside optical technologies intended to improve performance and compactness.

B. Key Parameters Trade-offs and Potential Solutions

  • As field of view increases, aberration correction becomes more difficult, making architecture selection important for image quality and broad coverage.Conventional designs include single-camera scanning and multi-camera stitching, each with constraints such as lack of real-time capture or calibration errors and added volume.Source: B. Key Parameters Trade-offs and… · p. 4
  • Table I compares different panoramic imaging system architectures, while Fig. 2 illustrates parameter trade-offs for a fisheye camera.Fig. 2 depicts inscribed-circle and circumscribed-circle imaging modes.Source: Figure 2 · p. 4, Table I · p. 5
  • Small-F-number and long-focal-length systems support higher illumination or high-resolution sensors, while infrared, visible, ultraviolet, and polarized-light systems extend sensing modalities.Zoom systems switch between large and small fields of view, multi-channel systems offer different views, and dichroic films or polarization techniques are described for eliminating central blind areas in catadioptric systems.Source: B. Key Parameters Trade-offs and… · p. 5
  • Freeform surfaces, thin-plate optics, and metasurfaces offer additional design freedom for improving image quality while reducing panoramic-system volume.The review links these technologies to high-precision manufacturing and potential applications including autonomous driving, medical treatment, and satellite navigation.Source: B. Key Parameters Trade-offs and… · p. 5

III. ARCHITECTURES AND PROPERTIES OF PANORAMIC IMAGING SYSTEMS

Field-of-view stitching combines views captured through single-camera scanning or multiple cameras to form large-field images. The approaches differ in capture conditions and calibration demands.

A. Single Camera Scanning and Multiple Cameras Stitching

  • Single-camera scanning rotates a camera on a high-precision mechanical stage to combine views into a large-field image.It suits relatively static scenes, requires accurate scanning, and cannot provide gaze imaging during the time needed to rotate and scan; one cited setup forms a 120° image.Source: A. Single Camera Scanning and Multiple… · p. 5
  • Multiple cameras can shoot simultaneously to produce a 360° surround view, but stitching depends on camera positioning, calibration, and image-stitching algorithms.Differences in camera settings and installation angles affect the stitching.Source: A. Single Camera Scanning and Multiple… · pp. 5–6
  • Multi-camera examples include a nine-camera low-cost long-wave infrared system and a four-camera visible-band portable stereo 360° design.The review also describes optimizing a two-mirror pyramid camera’s geometry and imager placement to maximize field of view, sensor utilization, and image uniformity.Source: A. Single Camera Scanning and Multiple… · p. 6

B. Fisheye Panoramic System

Fisheye systems provide single-camera panoramic views exceeding 180° but trade wide coverage against distortion and compactness; panomorph lenses redistribute image detail toward regions of interest.

B. Fisheye Panoramic System

  • A single fisheye camera avoids image stitching and typically covers more than 180°, with some systems reaching 280° without a central blind area.Its large front lenses and edge distortion remain design concerns.Source: B. Fisheye Panoramic System · p. 6
  • Fisheye systems use front negative meniscus lenses to compress wide-angle input before relay-lens aberration correction, but f-θ distortion can reach 15%–20%.Compactness, color correction, distortion, and off-axis aberrations also require attention.Source: B. Fisheye Panoramic System · pp. 6–7

C. Panomorph Imaging System

  • Panomorph lenses use cylindrical and toroidal surfaces to control anamorphic and distortion profiles, increasing sensor use and pixel density in regions of interest.The cited examples show a larger car image and more compressed edge content than with fisheye imaging.Source: C. Panomorph Imaging System · p. 7
  • A panomorph lens provides twice the border resolution of its center resolution, supporting surveillance, driving monitoring, vehicle positioning, and endoscopic procedures.The cited applications also include infrared positioning and urban security surveillance.Source: C. Panomorph Imaging System · pp. 7–8

D. Catadioptric Panoramic System

Catadioptric panoramic systems combine mirrors and refractive relay optics to capture surrounding views, with design trade-offs involving distortion, blind areas, and calibration. Their architectures support multispectral imaging, 3D sensing, endoscopy, and other applications.

D. Catadioptric Panoramic System

  • A front mirror reflects surrounding 360° light toward a refractive relay group, which corrects aberrations for imaging.Compared with fisheye optics, the architecture uses fewer front lenses, but its f-θ distortion is usually greater than 10%.Source: D. Catadioptric Panoramic System · p. 8
  • Single-mirror systems can leave a central blind area, while distorted checkerboard images can reduce camera-calibration accuracy; iterative refinement reduces feature-point reprojection error by 39%.Single-mirror designs may use aspheric surfaces, and one odd-aspheric system reports image distortion below 0.5%.Source: D. Catadioptric Panoramic System · p. 9
  • A two-mirror system with mirrors bent in the same direction reduces the blind-area ratio to less than 3.67%.The architecture is also described for low-F-number long-wave infrared systems; comparison of same- and opposite-direction mirror layouts is discussed.Source: D. Catadioptric Panoramic System · p. 10
  • Catadioptric designs support visible–ultraviolet and visible–infrared imaging, panoramic stereo and 3D reconstruction, thermal sensing, and robotic perception.Other applications described include endoscopic imaging, infrared periscopes, and ghost panoramic single-pixel imaging.Source: D. Catadioptric Panoramic System · pp. 9–10

E. Monocentric Panoramic System

Monocentric systems use spherical surfaces sharing a center of curvature, with fiber relays and multiscale designs supporting compact wide-field imaging. Hyper-hemispheric lenses extend panoramic coverage and also appear in star-tracker applications.

E. Monocentric Panoramic System

  • CurvACE combines a microlens array, neuromorphic photodetectors, and a flexible circuit board into 630 ommatidia with a 180°×60° field of view.A separate monocentric monitoring lens has a 140° field of view, 7.88 mm focal length, 14.47 mm total length, and F/1.5.Source: E. Monocentric Panoramic System · p. 10
  • A monocentric lens uses spherical optical surfaces sharing a center of curvature; fiber bundles can relay its curved image to flat sensors.A cited fiber-coupled prototype is F/1.35, 30 megapixels, and 126° FoV, and image processing improves its relayed image.Source: E. Monocentric Panoramic System · pp. 10–11
  • A Galilean monocentric multiscale system reduces total volume by 10 times relative to an earlier Keplerian design, while other work develops folding and computational-imaging approaches.A later analytical model reports better aberration performance for GMMS than KMMS; monocentric light-field systems combine the lens with light-field sensors.Source: E. Monocentric Panoramic System · pp. 11–12

F. Hyper-hemispheric Lens

  • Hyper-hemispheric designs span reported fields of view of 360°×260° and 360° azimuth by 270° zenith, while a panoramic channel can retain a central blind area.A front channel can use the central sensor area, and a semi-reflective film can provide continuous front and panoramic imaging.Source: F. Hyper-hemispheric Lens · p. 12
  • A hyper-hemispheric lens used for star tracking supports attitude estimation better than 1° accuracy and a success rate of about 98%.The method combines computer-vision and robotics concepts, including template matching and point-cloud registration.Source: F. Hyper-hemispheric Lens · p. 13

G. Panoramic Annular Lens

Panoramic annular lenses use a compact catadioptric block to map a 360° surround view onto an annular image, although the front mirror can create a central blind area. The review surveys design refinements and uses spanning imaging, sensing, and scene perception.

G. Panoramic Annular Lens

  • A panoramic annular lens combines an optical block, relay group, and sensor; two refractions and two reflections form an annular image with a central blind area.The system uses flat-cylinder projection, and the front small mirror produces the blind area.Source: G. Panoramic Annular Lens · p. 13
  • Panoramic annular systems support high-definition imaging, infrared monitoring, endoscopy, deformation monitoring, and light-field display.An ogive design reports f-θ distortion below 0.8% for a 360°×(40°∼95°) field of view and is described as suitable for SLAM.Source: G. Panoramic Annular Lens · pp. 13–14
  • Panoramic annular sensing extends to polarization-based pavement observation, multisensor fusion for autonomous driving, and panoramic atmospheric hyperspectral detection.The review also summarizes the development history and key parameters of panoramic architectures in Fig. 10 and Table II.Source: G. Panoramic Annular Lens · pp. 14–15
  • Designs address blind areas, distortion, and compactness with dichroic filters, polarizing films, freeform surfaces, and three spherical lenses.A cited compact system reports no blind area or vignetting, shared relay optics, and a 1.5 compact ratio; a three-lens system covers 360°×(25°∼100°).Source: G. Panoramic Annular Lens · p. 15
  • Dual-channel designs support coarse panoramic search and fine observation, zoom without shifting the target from the field-of-view center, and front-plus-surround observation.Other examples include a dual-view endoscope without a blind area and a 360°×230° system combining front and panoramic channels.Source: G. Panoramic Annular Lens · pp. 14–15

IV. NEW ENGINES OF PANORAMIC OPTICAL SYSTEMS

The review presents freeform optics as a route to compact, high-image-quality panoramic systems, with expanded design freedom enabling varied structures and functions. It also identifies design and manufacturing challenges that remain.

A. Freeform Surface

  • Freeform surfaces offer more design freedom than traditional spherical surfaces, improving imaging performance while reducing system volume.They are defined here as surfaces without axis-rotational or translation-rotational symmetry.Source: A. Freeform Surface · p. 15
  • Panoramic systems use rotationally symmetric aspheres and asymmetric freeforms, including Alvarez and XY-polynomial surfaces, for designs and zoom functions difficult with traditional lenses.Examples include odd, even, Q-type, extended-polynomial, annularly stitched, ogive, and even-ogive aspheres.Source: A. Freeform Surface · p. 15
  • Freeform design needs stronger global optimization, while fabrication, testing, and assembly still need greater accuracy and efficiency.The review identifies these as continuing challenges for the technology.Source: A. Freeform Surface · p. 15

B. Thin-plate Optics

Thin-plate optics combines thin Fresnel or diffractive elements with computational imaging to build lightweight, compact systems. The review describes image recovery mechanisms and outstanding design and performance challenges.

B. Thin-plate Optics

  • Thin-plate optics pairs Fresnel, diffractive, or other thin optical elements with computational imaging to recover or enhance images.The approach is intended to enable lightweight panoramic systems.Source: B. Thin-plate Optics · pp. 15–16
  • Computational imaging can encode and decode optical information and analyze multidimensional light-field information captured through scattering, polarization, and bionic technologies.Fresnel lenses reduce lens thickness while maintaining lens curvature; diffraction occurs when optical surface processing approaches the imaging wavelength.Source: B. Thin-plate Optics · p. 16
  • Diffraction-coded lenses create distinct point-spread functions for different spectral distributions, enabling computational phase-difference recovery and image reconstruction.The review describes this as one route to image restoration.Source: B. Thin-plate Optics · p. 18, Figure 11 · p. 18
  • Thin-plate optics still needs more numerically stable, faster-converging design algorithms, higher diffraction efficiency, and a broader usable wavelength range.The review notes that micro-nano processing has matured, but these challenges remain.Source: B. Thin-plate Optics · p. 18

C. Wide-FoV Metalens

The review describes wide-FoV metalenses as an ultrathin design option for panoramic systems, while noting unresolved optical and fabrication challenges. It also surveys panoramic semantic segmentation approaches for surrounding scene understanding, including methods addressing data scarcity, distortion, and efficient deployment.

C. Wide-FoV Metalens

  • Metalenses use arranged subwavelength scatterers to control optical phase, offering an ultrathin, lightweight alternative to conventional lenses.The review describes the beam as focusable to the diffraction limit and notes ultra-short focal lengths.Source: C. Wide-FoV Metalens · pp. 18–19
  • Metalens research targets large FoV, achromatic and broadband operation, and high spectral resolution, with potential for compact, high-performance panoramic systems.The review points to work on these properties and a comparison of wide-FoV metalenses.Source: C. Wide-FoV Metalens · p. 19, Table III · p. 19
  • Metalenses still face chromatic aberration, low focusing efficiency, and difficult data transfer and iteration in overall system optimization.The review identifies broad-spectrum metalenses as a future research focus.Source: C. Wide-FoV Metalens · p. 19
  • Metalens diameter must increase for integration with conventional optics; combining the technologies may support future hyper-hemispheric panoramic instruments.The review frames higher-performance miniature panoramic scene perception as a future possibility.Source: C. Wide-FoV Metalens · p. 19

A. Semantic Scene Understanding with Image Segmentation

  • Panoramic semantic segmentation supports pixel-level scene understanding, while 360° cameras can provide segmentation across the entire surrounding scene.The review surveys segmentation applications for robotics and autonomous driving; Fig. 12 illustrates several panoramic segmentation approaches.Source: A. Semantic Scene Understanding with… · p. 19, §V · p. 19, Figure 12 · p. 20
  • Wide-FoV segmentation faces scarce annotated data because pixel labeling is labor-intensive, especially for distorted, complex panoramic images with many small objects.Fisheye methods synthesize training data from pinhole datasets, and zoom or seven-DoF augmentation addresses variation in camera orientation, posture, and focal length.Source: A. Semantic Scene Understanding with… · p. 19
  • PASS reuses large-scale pinhole data for single-camera panoramic segmentation, using network adaptation and padding to produce seamless results across segments.Later variants add efficient models, attention connections, or multi-source omni-supervised learning; the review notes that segmenting panoramas can incur computation overhead and lose global context.Source: A. Semantic Scene Understanding with… · pp. 19–20
  • Domain adaptation and generalization transfer from labeled pinhole images toward panoramic segmentation, while panoptic methods add instance IDs to pixel-level semantic classes.Dense contrastive pretraining is reported to improve generalization on unseen panoramic images; other work addresses distortion and model efficiency.Source: A. Semantic Scene Understanding with… · pp. 20–21

B. Geometric Scene Understanding with Depth Estimation

The review surveys supervised and unsupervised approaches to dense panoramic depth estimation, including projection-aware methods that address distortion and use complementary representations. It also describes combined learning strategies and depth completion for panoramic imagery.

B. Geometric Scene Understanding with Depth Estimation

  • Supervised panoramic depth methods use ground-truth depth and include synthetic training data, distortion-aware convolutions, projection transformations, geometric priors, and feature-map compression.Some methods assume indoor panoramas align with gravity; others adapt convolution dilation or fuse equirectangular and cube-map representations.Source: B. Geometric Scene Understanding with… · p. 21, Figure 13 · p. 22
  • Unsupervised methods do not require ground-truth depth during training, reducing data-preparation burden; examples use spherical view synthesis or distortion-aware depth-image-based rendering.The review says unsupervised panoramic depth estimation has received less attention than supervised methods.Source: B. Geometric Scene Understanding with… · p. 21, Figure 13 · p. 22
  • Equirectangular projection provides full-FoV observations, while cube maps provide distortion-free representations for methods that fuse the two.The review also describes tangent-image patches and spherical-tangent-domain transformers as approaches to panoramic depth estimation.Source: B. Geometric Scene Understanding with… · p. 21
  • Panoramic depth completion targets the depth channel of a single 360° RGB-D pair, and depth estimation or completion may improve mobile-robot perception and operational efficiency in large-scale environments.The review also describes combining supervised and self-supervised learning to compensate for weaknesses of each approach.Source: B. Geometric Scene Understanding with… · pp. 21–22

C. Visual Localization, Odometry, and SLAM

The review surveys panoramic approaches to visual localization, odometry, and SLAM. Wider views and panoramic camera models are used to address appearance variation, rapid motion, and mapping challenges.

Panoramic visual localization

  • Panoramic localization methods address appearance and viewpoint variation through panorama matching, active deep descriptors, coarse-to-fine representations, and semantic cues,,,, [363],,.Prior work also finds that aggregating pinhole views can be preferable to a single panorama descriptor, and that front- and back-facing views are especially descriptive for long-term place recognition,.Source: C. Visual Localization, Odometry, and… · p. 22

Panoramic visual odometry

  • Panoramic visual odometry adapts initialization, tracking, and depth filtering to improve robustness to rapid motion and dynamic objects, while wide-FoV methods also report benefits and accuracy trade-offs across scene types,,,.Semantic-guided keypoint extraction and correspondence search further improved pose-estimation robustness in PALVO.Source: C. Visual Localization, Odometry, and… · pp. 22–23

Panoramic visual SLAM

  • Panoramic SLAM extends tracking, mapping, and loop closure to omnidirectional camera models, with reported gains in localization accuracy, robustness, or performance over narrow-FoV baselines,,,.Systems include semi-dense feature mapping, panoramic annular lenses, and filtering of features to the annular useful area; additional work covers single-camera mapping and sensor fusion,,,,, –, –.Source: C. Visual Localization, Odometry, and… · pp. 23–24, Figure 14 · p. 23

D. Optical Flow Estimation, Object Detection, and Others

The review covers panoramic optical flow, object detection, and other downstream vision tasks. Methods adapt perspective models, datasets, or network operations to panoramic geometry and its distortions.

Optical flow estimation

  • Panoramic flow estimation supports temporal scene cues, but large-scale panoramic flow datasets are scarce, motivating domain adaptation and panoramic-specific data generation,,,,,,.PanoFlow provides a cyclic framework adaptable to perspective learning models and the Flow360 dataset for training and quantitative analysis [396].Source: D. Optical Flow Estimation, Object… · pp. 24–25

Object detection

  • Panoramic object detection adapts detectors and annotations to equirectangular images, including distortion augmentation, anchor-position information, and EquiConvs for panoramic geometry,,,,.Some methods also recover 3D object position or jointly estimate 2D masks and 3D bounding boxes,.Source: D. Optical Flow Estimation, Object… · p. 25

Other downstream tasks

  • For indoor scene understanding, panoramic layout methods estimate room geometry, with later work recovering detailed object shape, pose, and location from a single panorama, [416],,,.Panoramic video surveillance work detects anomalous objects, while saliency methods predict human visual attention, including a weakly supervised model and Wild-360 annotations,,.Source: D. Optical Flow Estimation, Object… · p. 25

VI. CONCLUSIONS AND PERSPECTIVES

The review identifies emerging optical technologies and multidimensional sensing as directions for panoramic imaging, alongside challenges in design, fabrication, and practical learning. It anticipates compact, intelligent systems for broader scene understanding.

VI. CONCLUSIONS AND PERSPECTIVES

  • Freeform surfaces, thin-plate optics, and metalenses offer alternatives to traditional components and greater design freedom, supporting the prospect of compact, lightweight, high-performance panoramic systems [25].Ultra-precision manufacturing is described as support for micro-nano processing of these optical technologies [25].Source: §VI · p. 25
  • The review synthesizes panoramic optical systems and downstream perception applications, clarifying their advantages, progress, limitations, and future challenges [28].It anticipates applications including autonomous driving, medical health, security monitoring, exploration, and intelligent robots [35].Source: §VI · pp. 25–26
  • The review calls for multi-FoV and multifunctional systems, including zoom and polarization, with high temporal and spectral resolution for scene-understanding applications [31].It also points to combining panoramic optics with event sensors or active-sensing LiDAR for multimodal surrounding perception [33].Source: §VI · pp. 25–26
  • Fusion of new intelligent sensors with computational imaging is identified as a direction for developing panoramic instruments for scene understanding [32].The review anticipates combining emerging sensors, precision manufacturing, computational imaging, and AI in low-power, small-volume panoramic instruments [35].Source: §VI · p. 26
  • Because whole-panorama annotation is difficult and expensive, the review points to unsupervised or weakly supervised learning and adaptation from pinhole imagery as practical research directions [34].The review frames design and fabrication, including fitting high-performance systems into compact space, as continuing challenges [25].Source: §VI · pp. 25–26
Loading 2205.05570v2…