Small object detection - definition
Small object detection is the ability of a computer vision system to locate objects that occupy a limited number of pixels in an image or video frame. In image and video anonymization, the term is especially relevant to faces and license plates that are far from the camera, partially obscured, affected by motion blur, or recorded at low resolution.
A small face or license plate may still be visible to a human reviewer, but it can contain too little reliable visual information for an automated detector to identify it consistently. This creates a privacy risk: an automated anonymization workflow may fail to apply a blur to a distant face or license plate, even though the element remains identifiable or potentially identifiable in the source footage.
Small object detection is a technical computer vision problem, not a legal category. There is no universal pixel threshold at which a face or license plate becomes "small." The threshold depends on the detection model, input resolution, camera angle, compression level, lighting, object contrast, and the required recall rate.
Why distant faces and license plates require manual verification
Automated face detection and license plate detection depend on visual patterns. When an object occupies few pixels, the model has less information about its shape, edges, texture, and context. Downscaling during preprocessing can further reduce the object to a size at which detection becomes unreliable.
The Common Objects in Context (COCO) evaluation framework classifies a small object as one with an area below 322 pixels, or 1,024 pixels2. This benchmark definition is useful for comparing models, but it is not a privacy acceptance criterion. A face or license plate may require review even when its bounding-box area exceeds this value.
Manual verification is necessary because a missed detection is different from a false detection. A false detection may result in unnecessary blurring. A missed detection can leave a face or license plate visible in released footage.
- Distance: objects farther from the camera occupy fewer pixels.
- Compression: block artifacts and loss of fine detail can obscure facial contours and license plate characters.
- Motion: camera movement or object movement can produce directional blur.
- Occlusion: people, vehicles, vegetation, reflections, and shadows can conceal part of an object.
- Viewing angle: profile faces and angled license plates differ from the frontal examples used in many training datasets.
- Low contrast: poor lighting, glare, rain, or overexposure can reduce the separation between the object and its background.
Small object detection in image and video anonymization
For privacy protection, small object detection is used before a blur or mask is applied. The system first identifies a possible face or license plate, then creates a region of interest around it, and finally applies an anonymization effect to that region. Detection quality therefore directly affects anonymization completeness.
In video, verification must cover more than a single frame. A face or license plate can be visible only briefly, appear at the edge of the frame, or become small as a person or vehicle moves away from the camera. Object tracking can help associate detections across frames, but tracking does not reliably recover an object that was never detected.
Gallio PRO automatically detects and blurs faces and license plates in imported images and video files. It does not perform real-time anonymization or video stream anonymization. Where distant, small, partially visible, or uncertain objects remain, an operator should review the output and use the built-in manual editor to apply a blur where necessary. Gallio PRO does not automatically detect logos, tattoos, name badges, documents, monitor content, bodies, or silhouettes. Such elements can be blurred manually when they are relevant to the intended disclosure or privacy assessment.
Key parameters and metrics for small object detection
Detection quality should be measured on footage that resembles the intended production material. A model tested only on close, well-lit faces may not perform comparably on distant surveillance-style footage, vehicle recordings, or wide-angle camera views.
Parameter or metric | Meaning for anonymization
|
|---|---|
Input resolution | Higher resolution can preserve more visual information for small faces and license plates, but it increases processing time and memory use. |
Object pixel area | A practical indicator of whether the detector has enough visual detail. It should be measured before and after any resizing step. |
Recall | The proportion of relevant faces or license plates detected by the system. For privacy review, missed objects are the central concern. |
Precision | The proportion of detections that are correct. Low precision increases review work but is usually less privacy-critical than low recall. |
Intersection over Union (IoU) | The overlap between the predicted detection area and the reference area. Incomplete boxes can leave edges of a face or plate insufficiently blurred. |
Frame sampling and tracking | Controls whether brief appearances are examined and whether blur regions remain aligned across consecutive frames. |
Recall is calculated as follows:
Recall = true positives / (true positives + false negatives)
For a privacy workflow, evaluation should include difficult examples: distant pedestrians, parked vehicles in the background, subjects entering or leaving the frame, night footage, reflected faces, and license plates affected by glare. The review protocol should document the footage type, detector configuration, sampling method, reviewer decisions, and unresolved limitations.
Technical approaches and limitations
Modern object detectors often use multi-scale feature extraction. Feature Pyramid Networks, for example, combine image features from different network layers so that the model can process objects at multiple scales. Image tiling is another approach. It divides a high-resolution frame into smaller regions so that distant faces and license plates occupy more pixels during detection.
These methods can improve small object recall, but they do not remove the need for human review. Tiling may increase processing time and can create boundary problems when an object spans two tiles. Enlarging an image can make an object appear larger, but it cannot restore detail that was absent from the original recording. Automated detection results should therefore be treated as review candidates rather than proof that all visible faces and license plates have been anonymized.
Standards and references
The following sources provide technical context for evaluating small object detection and image quality. They do not establish a universal minimum size for privacy-safe face or license plate blurring.
- Tsung-Yi Lin et al., Microsoft COCO: Common Objects in Context, 2014. The COCO evaluation protocol defines small objects by area.
- Tsung-Yi Lin et al., Feature Pyramid Networks for Object Detection, 2017. This publication describes multi-scale feature pyramids for object detection.
- National Institute of Standards and Technology (NIST), Face Recognition Vendor Test - Face Image Quality. NIST explains how image quality factors can affect automated face analysis performance.
- ISO/IEC 29794-5:2010, Information technology - Biometric sample quality - Part 5: Face image data. This standard addresses quality properties of face image data.