Short-form video feeds now curate billions of clips for predominantly young audiences, yet the same algorithms that drive engagement often amplify sexualised portrayals—especially of women—in ways that can shape body image and social norms. This thesis introduces an end-to-end computer-vision pipeline that treats every video frame as three complementary signals: objects reveal context, body landmarks trace movement, and a coarse gender cue situates social meaning. Dedicated detection, pose, and appearance models are trained on hand-curated datasets and then fused into a single routine that annotates unseen TikTok-style videos while logging frame-level statistics. The system does not decide what is or is not “sexual”; instead, it flags recurrent visual combinations—mirrors with flexing poses, low-angle phone shots, revealing workout wear—that prior scholarship links to sexualised framing. Experiments show that this multimodal approach captures context more reliably than any single branch alone, though its scope is limited by binary gender labels, Western-skewed data, and the cultural fluidity of sexual meaning. The work offers both a practical tool for content analysis and a roadmap for more nuanced, ethically informed moderation research.
I feed video di breve durata ora selezionano miliardi di clip per un pubblico prevalentemente giovane, eppure gli stessi algoritmi che guidano il coinvolgimento spesso amplificano le rappresentazioni sessualizzate, soprattutto di donne, in modi che possono plasmare l'immagine corporea e le norme sociali. Questa tesi introduce una pipeline di visione artificiale end-to-end che tratta ogni fotogramma video come tre segnali complementari: gli oggetti rivelano il contesto, i punti di riferimento del corpo tracciano il movimento e un indizio di genere approssimativo colloca il significato sociale. Modelli dedicati di rilevamento, posa e aspetto vengono addestrati su set di dati curati manualmente e poi fusi in un'unica routine che annota video inediti in stile TikTok, registrando statistiche a livello di fotogramma. Il sistema non decide cosa sia o non sia "sessuale"; segnala invece combinazioni visive ricorrenti – specchi con pose flettenti, scatti con il telefono dal basso, abbigliamento sportivo succinto – che studi precedenti collegano a inquadrature sessualizzate. Gli esperimenti dimostrano che questo approccio multimodale cattura il contesto in modo più affidabile rispetto a qualsiasi singolo ramo, sebbene la sua portata sia limitata da etichette di genere binarie, dati orientati all'Occidente e dalla fluidità culturale del significato sessuale. Il lavoro offre sia uno strumento pratico per l'analisi dei contenuti sia una tabella di marcia per una ricerca sulla moderazione più articolata ed eticamente informata.
Understanding Online Sexualization Markers: A Tiktok Content Analysis
BAWM, ROSE MARY
2024/2025
Abstract
Short-form video feeds now curate billions of clips for predominantly young audiences, yet the same algorithms that drive engagement often amplify sexualised portrayals—especially of women—in ways that can shape body image and social norms. This thesis introduces an end-to-end computer-vision pipeline that treats every video frame as three complementary signals: objects reveal context, body landmarks trace movement, and a coarse gender cue situates social meaning. Dedicated detection, pose, and appearance models are trained on hand-curated datasets and then fused into a single routine that annotates unseen TikTok-style videos while logging frame-level statistics. The system does not decide what is or is not “sexual”; instead, it flags recurrent visual combinations—mirrors with flexing poses, low-angle phone shots, revealing workout wear—that prior scholarship links to sexualised framing. Experiments show that this multimodal approach captures context more reliably than any single branch alone, though its scope is limited by binary gender labels, Western-skewed data, and the cultural fluidity of sexual meaning. The work offers both a practical tool for content analysis and a roadmap for more nuanced, ethically informed moderation research.| File | Dimensione | Formato | |
|---|---|---|---|
|
Bawm_Rose_Mary.pdf
Accesso riservato
Dimensione
4.12 MB
Formato
Adobe PDF
|
4.12 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/86895