The conditions of visibility : threshold collapse and the production of subjects in computer vision
الملخص
Aerial surveillance systems increasingly rely on computer vision to identify human presence in drone imagery, supporting forms of situational awareness used in security and monitoring contexts. This report examines how computer vision systems render human bodies visible within aerial imagery. It analyzes how visibility is produced through model architecture, data conditions, and decision thresholds. Using a YOLO-based detector applied to the VisDrone dataset, the report introduces controlled corruptions including noise, blur, darkening, and downsampling, along with variations in confidence thresholds. The results show that detection does not degrade smoothly. It exhibits threshold-like behavior where recall remains stable up to a point and then collapses abruptly. This collapse is not only caused by degraded input. Even when the model assigns non-zero confidence to potential detections, many are excluded by thresholding, creating a gap between internal representation and final output. As a result, bodies may be partially recognized yet absent in visible detections. These findings suggest that object detection operates as a system of conditional visibility. Visibility is produced through transformations and filtering operations that determine what counts as detectable. Detection systems establish the conditions under which bodies can appear, positioning object detection as a visual apparatus that constructs the subjects it identifies.
الكلمات المفتاحية
غير متوفر
التخصصات
رسائل وأطاريح مشابهة
- Deep Learning-Based Vision System for Intelligent Object Detection and Robotic Sorting Solution Anastasiia Stepanova · Tampere University · 2026
- Tracking an object with a depth-vision camera Josefína VLK · 2025
- Disease detection and mapping in olive groves using UAVs and deep learning for precision agriculture Maurício Morais · 2025
- Deep Learning Based Perception on Radar Data for Autonomous Driving Meyer, Michael · Christian-Albrechts-Universität zu Kiel · 2025
- Leveraging deep learning for visual understanding of videos Clement Xian Ren Tan · Nanyang Technological University · 2024
- Computer Vision for Automated Robotic Device Disassembly: Object Detection, Pose Estimation, and Action Prediction Ruiz, Sebastian · Georg-August-Universität Göttingen · 2024
