360-degree video anonymization - definition
360-degree video anonymization is the process of locating and redacting identifiable visual information in immersive video that captures a complete or near-complete scene around the camera. In practice, the main automated targets are faces and license plates. The process must account for the geometric distortion introduced by fisheye lenses and by 360-degree projection formats.
Unlike conventional flat video, a 360-degree recording may be stored as an equirectangular panorama, a cubemap, dual-fisheye frames, or another projection. A face near the edge of an equirectangular frame can be stretched horizontally. A face near the outer area of a fisheye lens can be curved, enlarged, compressed, or partly obscured by the lens boundary. These effects can reduce face detection accuracy and can cause an insufficiently positioned blur mask.
Effective anonymization therefore requires both object detection and geometry-aware redaction. The system must detect a face or license plate in the source representation, map the detection correctly across frames, and apply a mask that continues to cover the target despite movement, projection changes, and stitching artifacts.
Why lens geometry affects face detection
Fisheye lenses use an ultra-wide field of view. They do not preserve the straight-line geometry assumed by many computer-vision models trained primarily on perspective photographs. The relationship between a 3D scene point and its image position depends on the lens projection model.
Common fisheye projection models include the following:
- Equidistant projection: image radius is proportional to the viewing angle.
- Equisolid-angle projection: image radius is related to the sine of half the viewing angle.
- Stereographic projection: preserves angles locally but expands objects toward the image edge.
- Orthographic projection: projects the visible hemisphere onto a disk.
In a dual-fisheye recording, each lens captures approximately one hemisphere. The hemispheres are then stitched into a panorama. Faces and license plates near the stitch line may be split between lens images or may appear warped after stitching. A detector trained on ordinary perspective images may miss these targets or return an inaccurate bounding box.
Projection formats and redaction placement
The projection used for storage and processing determines how a blur mask should be calculated. A rectangular mask that is suitable for a standard camera frame may be unsuitable for an equirectangular panorama, particularly close to the poles or the left and right seam.
Representation | Typical geometric issue | Redaction consideration
|
|---|---|---|
Equirectangular panorama | Horizontal stretching near the top and bottom of the image; seam at 0°/360° longitude | Allow masks to wrap across the left and right image edges and enlarge masks near polar regions. |
Cubemap | Objects can cross cube-face boundaries | Maintain mask continuity across adjacent cube faces. |
Dual-fisheye source | Strong radial distortion and overlap between lenses | Detect before or after stitching, then validate coverage at lens and stitch boundaries. |
Perspective viewport | Only part of the sphere is visible at one time | Do not limit processing to a single viewer direction. Analyze the full recorded sphere. |
A robust workflow may map each detection to spherical coordinates rather than treating the source frame as a flat image. For an equirectangular frame with width W and height H, pixel coordinates can be converted approximately as follows:
longitude = 2π(x / W - 0.5)
latitude = π(0.5 - y / H)
This mapping helps preserve the location of a target when the footage is converted between equirectangular, cubemap, and viewport representations.
Detection, tracking, and mask propagation
360-degree video anonymization normally combines face detection or license plate detection with object tracking and temporal mask propagation. Detection identifies candidate targets. Tracking estimates where the target appears in adjacent frames. The redaction mask is then updated to maintain coverage throughout the target's visible duration.
Key technical controls include:
- Detection recall: the percentage of visible faces or license plates that are detected. Low recall creates unredacted exposure risk.
- Precision: the percentage of detections that are actual targets. Low precision can create unnecessary blurring.
- Intersection over Union (IoU): overlap between a predicted region and a ground-truth region. IoU alone may be insufficient for fisheye content because a rectangular region can include distorted background areas.
- Mask margin: additional pixels or angular distance around a detected target. A margin helps cover detector localization error, motion blur, and tracking drift.
- Temporal continuity: the absence of uncovered frames while a target remains visible.
Quality assurance should include manual review of difficult areas: lens peripheries, stitch seams, low-light frames, rapid camera movement, occluded faces, reflections, and targets that enter or leave the frame. Review should be performed on the final rendered output, not only on detector annotations.
Practical processing approach
A common approach is to create multiple perspective views from the spherical source. Each view has less distortion than the full equirectangular image and can be processed by a detector trained on conventional images. Detections from overlapping views are then transformed back to the original spherical coordinate system and merged.
This approach increases computational cost because the same target can appear in multiple viewports. It also requires deduplication. A processing team should define viewport resolution, field of view, overlap, detection threshold, tracking tolerance, and blur-mask margin before production processing begins.
Gallio PRO supports automated face detection and license plate detection for recorded images and video. It does not perform real-time anonymization or video stream anonymization. Elements such as logos, tattoos, name badges, documents, and monitor content require manual redaction with the built-in editor when they must be obscured.
Challenges and limitations of 360-degree video anonymization
Geometric correction improves detection conditions but does not guarantee complete anonymization. A face may remain identifiable through partial visibility, unusual camera angles, reflection, motion blur, or a sequence of frames in which tracking fails. A license plate may be unreadable in one projection but legible after another viewer orientation or image enhancement.
Blur masks should therefore be evaluated at the output resolution and in the viewing formats that recipients can access. For immersive media, this includes panning across the scene, inspecting seam regions, and testing multiple viewer orientations.
Standards and references
Technical implementations should document the source projection, transformations, codec settings, detection model version, confidence thresholds, and review results. This information supports repeatability and auditability without retaining logs that contain face or license plate detections.
- ISO/IEC 23090-2:2021, Information technology - Coded representation of immersive media - Part 2: Omnidirectional Media Format, International Organization for Standardization, 2021.
- RFC 7946, The GeoJSON Format, Internet Engineering Task Force, 2016. Its longitude and latitude concepts are useful when documenting spherical target locations.
- Face Recognition Vendor Test (FRVT), National Institute of Standards and Technology. The program provides evaluation context for image-based face analysis systems, although deployment testing remains necessary for fisheye and panoramic footage.