Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.804048
Title: Adapting computer vision models to limitations on input dimensionality and model complexity
Author: Abbas, Alhabib
Awarding Body: UCL (University College London)
Current Institution: University College London (University of London)
Date of Award: 2020
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Abstract:
When considering instances of distributed systems where visual sensors communicate with remote predictive models, data traffic is limited to the capacity of communication channels, and hardware limits the processing of collected data prior to transmission. We study novel methods of adapting visual inference to limitations on complexity and data availability at test time, wherever the aforementioned limitations exist. Our contributions detailed in this thesis consider both task-specific and task-generic approaches to reducing the data requirement for inference, and evaluate our proposed methods on a wide range of computer vision tasks. This thesis makes four distinct contributions: (i) We investigate multi-class action classification via two-stream convolutional neural networks that directly ingest information extracted from compressed video bitstreams. We show that selective access to macroblock motion vector information provides a good low-dimensional approximation of the underlying optical flow in visual sequences. (ii) We devise a bitstream cropping method by which AVC/H.264 and H.265 bitstreams are reduced to the minimum amount of necessary elements for optical flow extraction, while maintaining compliance with codec standards. We additionally study the effect of codec rate-quality control on the sparsity and noise incurred on optical flow derived from resulting bitstreams, and do so for multiple coding standards. (iii) We demonstrate degrees of variability in the amount of data required for action classification, and leverage this to reduce the dimensionality of input volumes by inferring the required temporal extent for accurate classification prior to processing via learnable machines. (iv) We extend the Mixtures-of-Experts (MoE) paradigm to adapt the data cost of inference for any set of constituent experts. We postulate that the minimum acceptable data cost of inference varies for different input space partitions, and consider mixtures where each expert is designed to meet a different set of constraints on input dimensionality. To take advantage of the flexibility of such mixtures in processing different input representations and modalities, we train biased gating functions such that experts requiring less information to make their inferences are favoured to others. We finally note that, our proposed data utility optimization solutions include a learnable component which considers specified priorities on the amount of information to be used prior to inference, and can be realized for any combination of tasks, modalities, and constraints on available data.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.804048  DOI: Not available
Share: