Matteo Di Iorio
Multi-Scale Temporal Mamba for Dense Action Detection in Activities of Daily Living: Architectural Extensions and the Limits of Temporal Modelling.
Rel. Fabrizio Lamberti, Francois Bremond. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2026
Abstract
Temporal Action Detection (TAD) in untrimmed videos of Activities of Daily Living (ADL) requires models to process long-duration inputs, to capture temporal variations in actions, and to detect dense overlapping actions simultaneously. MS-Temba addresses this by extending Mamba to TAD with dilated State Space Models (SSMs). Each Temba block learns representations at a distinct temporal scale, and a lightweight Multi-scale Mamba Fuser aggregates them into boundary-accurate per-frame predictions. This thesis examines a deficit that the published results exhibit. Under action-conditioned metrics, MS-Temba scores roughly thirteen points lower at zero temporal tolerance than at wider ones. The thesis investigates where this deficit originates and whether temporal modelling can reach it at all.
First MS-Temba is reproduced with the released implementation, exposing a scheduler defect which disables cosine annealing throughout training inflating any result obtained with it
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Informazioni aggiuntive
Corso di laurea
Classe di laurea
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
