{"id":{"repo_id":"mcmaster","oai_identifier":"oai:macsphere.mcmaster.ca:11375/33187"},"canonical_url":"https://search.ndltd.org/etd/mcmaster/oai:macsphere.mcmaster.ca:11375/33187","repository":{"repo_id":"mcmaster","name":"McMaster University","base_url":"https://macsphere.mcmaster.ca/server/oai/request"},"display":{"title":"Equivariant Sensor Fusion for Autonomous Driving Object Detection","abstract":"Geometric data augmentation is widely used in autonomous driving 3D object detection to improve robustness to transformations such as rotation, translation, and scaling. While effective, this approach increases training complexity and does not guarantee that the detector will respond predictably to transformations outside of those explicitly modelled. In this thesis, we investigate whether some of this robustness can instead be built directly into a BEVFusion-style detector through equivariance. To study this, we introduce two architectures. The first, EquiFusion, applies equivariant structure throughout the full detection pipeline. The second, HybridFusion, combines an equivariant stream with a parallel non-equivariant stream in order to balance geometric structure with feature flexibility. We also evaluate the effects of representational depth, non-equivariant stream placement, balanced resampling, and geometric data augmentation on the nuScenes benchmark. The results show that full equivariance alone does not work well. Across both augmentation and no-augmentation settings, EquiFusion performs consistently worse than the BEVFusion baseline, indicating that the fully equivariant constraint is too restrictive when imposed throughout the model. In contrast, HybridFusion performs much more strongly, especially in the no-augmentation setting. Most notably, unaugmented HybridFusion outperforms augmented BEVFusion in overall NDS. These results show that equivariance can be useful in BEVFusion-style 3D object detection, but that its strongest role is not as a full replacement for non-equivariant processing. Instead, the best performance is obtained when equivariant and non-equivariant features are combined within a hybrid architecture.","abstract_html":"Geometric data augmentation is widely used in autonomous driving 3D object detection to improve robustness to transformations such as rotation, translation, and scaling. While effective, this approach increases training complexity and does not guarantee that the detector will respond predictably to transformations outside of those explicitly modelled. In this thesis, we investigate whether some of this robustness can instead be built directly into a BEVFusion-style detector through equivariance. To study this, we introduce two architectures. The first, EquiFusion, applies equivariant structure throughout the full detection pipeline. The second, HybridFusion, combines an equivariant stream with a parallel non-equivariant stream in order to balance geometric structure with feature flexibility. We also evaluate the effects of representational depth, non-equivariant stream placement, balanced resampling, and geometric data augmentation on the nuScenes benchmark. The results show that full equivariance alone does not work well. Across both augmentation and no-augmentation settings, EquiFusion performs consistently worse than the BEVFusion baseline, indicating that the fully equivariant constraint is too restrictive when imposed throughout the model. In contrast, HybridFusion performs much more strongly, especially in the no-augmentation setting. Most notably, unaugmented HybridFusion outperforms augmented BEVFusion in overall NDS. These results show that equivariance can be useful in BEVFusion-style 3D object detection, but that its strongest role is not as a full replacement for non-equivariant processing. Instead, the best performance is obtained when equivariant and non-equivariant features are combined within a hybrid architecture.","abstract_has_math":false,"creators":["Grandhi, Alaap"],"institution":null,"degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":"Computing and Software","school":null,"contributors":[],"advisors":["Giamou, Matthew"],"committee_chairs":[],"committee_members":[],"year":2026,"date_issued":"2026","date_published":"2026","updated_at":"2026-11-18T19:32:33Z","subjects":["Geometric Deep Learning","Deep Learning","Machine Learning","Computer Vision","Autonomous Driving","Equivariant Machine Learning","Sensor Fusion","Object Detection"],"languages":[],"rights":["Attribution 2.5 Canada"],"rights_urls":["http://creativecommons.org/licenses/by/2.5/ca/"],"identifier_entries":[{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://doi.org/10.71548/50"],"render_values":[{"text":"https://doi.org/10.71548/50","href":"https://doi.org/10.71548/50","code":true}]}],"person_identifier_entries":[]},"links":{"outbound_url":"https://hdl.handle.net/11375/33187","outbound_label":"Handle","outbound_source":"dc:identifier.uri","record_url":"https://hdl.handle.net/11375/33187","record_label":"Repository record","record_kind":"handle","record_source":"dc:identifier.uri","download_url":null,"download_label":null,"download_source":null,"access_status":null,"embargo_until":null,"access_note":null},"source_record":{"url":"https://macsphere.mcmaster.ca/server/oai/request?verb=GetRecord&metadataPrefix=dim&identifier=oai%3Amacsphere.mcmaster.ca%3A11375%2F33187","prefix":"dim"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["Giamou, Matthew"]},{"key":"dc:contributor.department","label":"Department","values":["Computing and Software"]},{"key":"dc:creator","label":"Author","values":["Grandhi, Alaap"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.accessioned","label":"Dc Date Accessioned","values":["2026-06-01T20:33:03Z"]},{"key":"dc:date.issued","label":"Date","values":["2026"]},{"key":"dc:type","label":"Dc Type","values":["Thesis"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Geometric Deep Learning","Deep Learning","Machine Learning","Computer Vision","Autonomous Driving","Equivariant Machine Learning","Sensor Fusion","Object Detection"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:rights","label":"Dc Rights","values":["Attribution 2.5 Canada"]},{"key":"dc:rights.uri","label":"Rights URI","values":["http://creativecommons.org/licenses/by/2.5/ca/"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://hdl.handle.net/11375/33187","https://doi.org/10.71548/50"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["Geometric data augmentation is widely used in autonomous driving 3D object detection to improve robustness to transformations such as rotation, translation, and scaling. While effective, this approach increases training complexity and does not guarantee that the detector will respond predictably to transformations outside of those explicitly modelled. In this thesis, we investigate whether some of this robustness can instead be built directly into a BEVFusion-style detector through equivariance. To study this, we introduce two architectures. The first, EquiFusion, applies equivariant structure throughout the full detection pipeline. The second, HybridFusion, combines an equivariant stream with a parallel non-equivariant stream in order to balance geometric structure with feature flexibility. We also evaluate the effects of representational depth, non-equivariant stream placement, balanced resampling, and geometric data augmentation on the nuScenes benchmark. The results show that full equivariance alone does not work well. Across both augmentation and no-augmentation settings, EquiFusion performs consistently worse than the BEVFusion baseline, indicating that the fully equivariant constraint is too restrictive when imposed throughout the model. In contrast, HybridFusion performs much more strongly, especially in the no-augmentation setting. Most notably, unaugmented HybridFusion outperforms augmented BEVFusion in overall NDS. These results show that equivariance can be useful in BEVFusion-style 3D object detection, but that its strongest role is not as a full replacement for non-equivariant processing. Instead, the best performance is obtained when equivariant and non-equivariant features are combined within a hybrid architecture."]},{"key":"dc:description.degree","label":"Dc Description Degree","values":["Master of Science (MSc)"]},{"key":"dc:title","label":"Title","values":["Equivariant Sensor Fusion for Autonomous Driving Object Detection"]},{"key":"others:access_status","label":"Others Access Status","values":["open.access"]}]}],"canonical_facts":{"dc:contributor.advisor":["Giamou, Matthew"],"dc:contributor.department":["Computing and Software"],"dc:creator":["Grandhi, Alaap"],"dc:date.accessioned":["2026-06-01T20:33:03Z"],"dc:date.issued":["2026"],"dc:description.abstract":["Geometric data augmentation is widely used in autonomous driving 3D object detection to improve robustness to transformations such as rotation, translation, and scaling. While effective, this approach increases training complexity and does not guarantee that the detector will respond predictably to transformations outside of those explicitly modelled. In this thesis, we investigate whether some of this robustness can instead be built directly into a BEVFusion-style detector through equivariance. To study this, we introduce two architectures. The first, EquiFusion, applies equivariant structure throughout the full detection pipeline. The second, HybridFusion, combines an equivariant stream with a parallel non-equivariant stream in order to balance geometric structure with feature flexibility. We also evaluate the effects of representational depth, non-equivariant stream placement, balanced resampling, and geometric data augmentation on the nuScenes benchmark. The results show that full equivariance alone does not work well. Across both augmentation and no-augmentation settings, EquiFusion performs consistently worse than the BEVFusion baseline, indicating that the fully equivariant constraint is too restrictive when imposed throughout the model. In contrast, HybridFusion performs much more strongly, especially in the no-augmentation setting. Most notably, unaugmented HybridFusion outperforms augmented BEVFusion in overall NDS. These results show that equivariance can be useful in BEVFusion-style 3D object detection, but that its strongest role is not as a full replacement for non-equivariant processing. Instead, the best performance is obtained when equivariant and non-equivariant features are combined within a hybrid architecture."],"dc:description.degree":["Master of Science (MSc)"],"dc:identifier.uri":["https://hdl.handle.net/11375/33187","https://doi.org/10.71548/50"],"dc:rights":["Attribution 2.5 Canada"],"dc:rights.uri":["http://creativecommons.org/licenses/by/2.5/ca/"],"dc:subject":["Geometric Deep Learning","Deep Learning","Machine Learning","Computer Vision","Autonomous Driving","Equivariant Machine Learning","Sensor Fusion","Object Detection"],"dc:title":["Equivariant Sensor Fusion for Autonomous Driving Object Detection"],"dc:type":["Thesis"],"others:access_status":["open.access"]},"updated_at":"2026-11-18T19:32:33Z"}