Xiangbo Gong
Artist Similarity Assessment Using Instagram Video Content.
Rel. Cristina Emma Margherita Rottondi, Massimiliano Zanoni, Andrea Bianco. Politecnico di Torino, Corso di laurea magistrale in Data Science And Engineering, 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (23MB) | Preview |
|
|
Archive (ZIP) (Documenti_allegati)
- Altro
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (34MB) |
Abstract
This thesis examines whether visual information from artist-related videos can support artist similarity learning. It addresses the video branch of a broader multimodal artist-similarity project and evaluates this modality independently before combining it with audio, captions, or static images. Frame-level Contrastive Language-Image Pre-training (CLIP) embeddings provide the visual input. Each artist is represented by up to ten videos, with thirty sampled frames per video and a 768-dimensional CLIP embedding per frame. The main model, TripletNet1, preserves this structure through a hierarchical Transformer. Its first level summarises the frames within each video; its second aggregates the resulting video tokens into one artist representation.
Training uses triplet loss, cosine distance, and similarity relations derived from AllMusic and Spotify
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
