Parametric active learning techniques for 3D hand pose estimation
File(s)
Author(s)
Caramalau, Razvan
Type
Thesis
Abstract
Active learning (AL) has recently gained popularity for deep learning (DL) models due to efficient and informative sampling, especially when the models
require large-scale datasets. The DL models designed for 3D-HPE demand
accurate and diverse large-scale datasets that are time-consuming, costly and
require experts. This thesis aims to explore AL primarily for the 3D hand
pose estimation (3D-HPE) task for the first time.
The thesis delves directly into an AL methodology customised for 3D-HPE learners to address this. Because predominantly the learners are regression-based algorithms, a Bayesian approximation of a DL architecture is presented to model uncertainties. This approximation generates data and model-
dependent uncertainties that are further combined with the data representativeness AL function, CoreSet, for sampling. Despite being the first work, it
creates informative samples and minimal joint errors with less training data
on three well-known depth datasets.
The second AL algorithm continues to improve the selection following a
new trend of parametric samplers. Precisely, this is proceeded task-agnostic with a Graph Convolutional Network (GCN) to offer higher order of representations between labelled and unlabelled data. The newly selected unlabelled
images are ranked based on uncertainty or GCN feature distribution.
Another novel sampler extends this idea, and tackles encountered AL issues,
like cold-start and distribution shift, by training in a self-supervised way with
contrastive learning. It shows leveraging the visual concepts from labelled
and unlabelled images while attaining state-of-the-art results.
The last part of the thesis brings prior AL insights and achievements in a
unified parametric-based sampler proposal for the multi-modal 3D-HPE task.
This sampler trains multi-variational auto-encoders to align the modalities
and provide better selection representation. Several query functions are
studied to open a new direction in deep AL sampling.
require large-scale datasets. The DL models designed for 3D-HPE demand
accurate and diverse large-scale datasets that are time-consuming, costly and
require experts. This thesis aims to explore AL primarily for the 3D hand
pose estimation (3D-HPE) task for the first time.
The thesis delves directly into an AL methodology customised for 3D-HPE learners to address this. Because predominantly the learners are regression-based algorithms, a Bayesian approximation of a DL architecture is presented to model uncertainties. This approximation generates data and model-
dependent uncertainties that are further combined with the data representativeness AL function, CoreSet, for sampling. Despite being the first work, it
creates informative samples and minimal joint errors with less training data
on three well-known depth datasets.
The second AL algorithm continues to improve the selection following a
new trend of parametric samplers. Precisely, this is proceeded task-agnostic with a Graph Convolutional Network (GCN) to offer higher order of representations between labelled and unlabelled data. The newly selected unlabelled
images are ranked based on uncertainty or GCN feature distribution.
Another novel sampler extends this idea, and tackles encountered AL issues,
like cold-start and distribution shift, by training in a self-supervised way with
contrastive learning. It shows leveraging the visual concepts from labelled
and unlabelled images while attaining state-of-the-art results.
The last part of the thesis brings prior AL insights and achievements in a
unified parametric-based sampler proposal for the multi-modal 3D-HPE task.
This sampler trains multi-variational auto-encoders to align the modalities
and provide better selection representation. Several query functions are
studied to open a new direction in deep AL sampling.
Version
Open Access
Date Issued
2022-11
Date Awarded
2023-07
Copyright Statement
Creative Commons Attribution NonCommercial NoDerivatives Licence
Advisor
Kim, Tae-Kyun
Sponsor
Huawei (Firm)
Publisher Department
Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)