Chadha, Aaron; (2019) From Pixels to Spikes: Efficient Multimodal Learning in the Presence of Domain Shift. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Main_final.pdf - Accepted Version Download (8MB) | Preview

Abstract

Computer vision aims to provide computers with a conceptual understanding of images or video by learning a high-level representation. This representation is typically derived from the pixel domain (i.e., RGB channels) for tasks such as image classification or action recognition. In this thesis, we explore how RGB inputs can either be pre-processed or supplemented with other compressed visual modalities, in order to improve the accuracy-complexity tradeoff for various computer vision tasks. Beginning with RGB-domain data only, we propose a multi-level, Voronoi based spatial partitioning of images, which are individually processed by a convolutional neural network (CNN), to improve the scale invariance of the embedding. We combine this with a novel and efficient approach for optimal bit allocation within the quantized cell representations. We evaluate this proposal on the content-based image retrieval task, which constitutes finding similar images in a dataset to a given query. We then move to the more challenging domain of action recognition, where a video sequence is classified according to its constituent action. In this case, we demonstrate how the RGB modality can be supplemented with a flow modality, comprising motion vectors extracted directly from the video codec. The motion vectors (MVs) are used both as input to a CNN and as an activity sensor for providing selective macroblock (MB) decoding of RGB frames instead of full-frame decoding. We independently train two CNNs on RGB and MV correspondences and then fuse their scores during inference, demonstrating faster end-to-end processing and competitive classification accuracy to recent work. In order to explore the use of more efficient sensing modalities, we replace the MV stream with a neuromorphic vision sensing (NVS) stream for action recognition. NVS hardware mimics the biological retina and operates with substantially lower power and at significantly higher sampling rates than conventional active pixel sensing (APS) cameras. Due to the lack of training data in this domain, we generate emulated NVS frames directly from consecutive RGB frames and use these to train a teacher-student framework that additionally leverages on the abundance of optical flow training data. In the final part of this thesis, we introduce a novel unsupervised domain adaptation method for further minimizing the domain shift between emulated (source) and real (target) NVS data domains.

Download activity - last month

Export as JSONCSVXML

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item