Improving 3D Monocular Object Detection with Dynamic Loss Function Adjustments

dc.contributor.authorEvain, Alexandre
dc.contributor.authorKhemmar, Redouane
dc.contributor.authorOrzalesi, Mathieu
dc.contributor.authorAhmedali, Sofiane
dc.contributor.editorSkala, Václav
dc.date.accessioned2025-07-30T09:45:29Z
dc.date.available2025-07-30T09:45:29Z
dc.date.issued2025
dc.description.abstract-translatedIn 3D monocular object detection, optimizing the loss function is crucial for balancing multiple competing metrics, such as depth estimation, orientation, and object dimensions. Traditional approaches use a weighted sum of individual losses, allowing metric prioritization but risking training instability due to competition between terms. To address this, we first experimented with different loss function configurations to see how different loss interactions could emphasize specific metrics. These initial results demonstrated that abrupt changes in loss functions cause significant precision drops, therefore we decided to try dynamic loss functions adjustment, using transition functions to gradually shift metric emphasis over the training process. Among the tested transition functions, the Smoothstep function had the best balance across all metrics, followed by the Linear function, while the Smootherstep function provided strong initial performance but was eventually outperformed. Our results suggest that controlled, smooth transitions between different loss functions can enhance training stability and final detection accuracy, providing a way to improve 3D object detection models without overhauling their architectureen
dc.description.sponsorshipThis research is funded and supported by SEGULA Technologies. We would like to thank SEGULA Technologies for their collaboration and for allowing us to conduct this research. We would like to thank also the engineers of the Autonomous Navigation Laboratory (ANL) of IRSEEM for their support. In addition, this work was performed, in part, on computing resources provided by CRIANN (Centre Regional Informatique et d’Applications Numeriques de Normandie, Normandy, France).
dc.format8 s.cs
dc.format.mimetypeapplication/pdf
dc.identifier.doihttp://www.doi.org/10.24132/CSRN.2025-17
dc.identifier.issn2464-4617 (Print)
dc.identifier.issn2464-4625 (online)
dc.identifier.urihttp://hdl.handle.net/11025/62223
dc.language.isoenen
dc.publisherVaclav Skala - UNION Agencyen
dc.rights© Vaclav Skala - UNION Agencyen
dc.rights.accessopenAccessen
dc.subjectmonokulární 3D detekce objektůcs
dc.subjectoptimalizace ztrátové funkcecs
dc.subjectdynamické vyrovnávání ztrátcs
dc.subjectpřechodové funkcecs
dc.subjecthluboké učenícs
dc.subjectpočítačové viděnícs
dc.subjecttrénovací stabilitacs
dc.subject.translatedmonocular 3D object detectionen
dc.subject.translatedloss function optimizationen
dc.subject.translateddynamic loss adjustmenten
dc.subject.translatedtransition functionsen
dc.subject.translateddeep learningen
dc.subject.translatedcomputer visionen
dc.subject.translatedtraining stabilityen
dc.titleImproving 3D Monocular Object Detection with Dynamic Loss Function Adjustmentsen
dc.typekonferenční příspěvekcs
dc.typeconferenceObjecten
dc.type.statusPeer revieweden
dc.type.versionpublishedVersionen
local.files.count1*
local.files.size2193575*
local.has.filesyes*

Files

Original bundle
Showing 1 - 1 out of 1 results
No Thumbnail Available
Name:
C02.pdf
Size:
2.09 MB
Format:
Adobe Portable Document Format
License bundle
Showing 1 - 1 out of 1 results
No Thumbnail Available
Name:
license.txt
Size:
1.71 KB
Format:
Item-specific license agreed upon to submission
Description: