Adaptive Transfer Learning: a simple but effective transfer learning

Lee, Jung H.; Kvinge, Henry; Howland, Scott; New, Zachary; Buckheit, John; Phillips, Lauren A.; Skomski, Elliott; Hibler, Jessica; Corley, Courtney D.; Hodas, Nathan O.

doi:10.48550/arxiv.2111.10937

Cited by 1 publication

(3 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Training models using adult datasets have proven advantageous due to the substantial knowledge transfer from a larger pool of relevant data [17]. Transfer learning serves as a crucial mechanism for refining the model's weights to facilitate adaptation with newborns' smaller joint ratios.…”

Section: Discussionmentioning

confidence: 99%

“…Subsequently, the model undergoes specific training using a self-collected dataset focused on infants' postures. This sequential training approach aims to transfer knowledge from the adult dataset to the infant dataset, enhancing the model's ability to accurately detect and track the joints of newborns [17].…”

Section: Methodsmentioning

confidence: 99%

“…The preservation of spatial resolution is crucial to avoiding a significant reduction in accuracy when the layers remain the same until the deepest layer matches the output resolution. To address this issue, a convolutional accumulation [17] is incorporated and changed in the backbone block to save spatial resolution and reuse essential weights. By doing so, the number of participants required for the task can be minimized.…”

Section: Feature Extractormentioning

confidence: 99%

See 2 more Smart Citations

Two-Dimensional Ultra Light-Weight Infant Pose Estimation with Single Branch Network

Nguyen,

Nguyen-Quang,

Nguyen

et al. 2024

Applied Sciences

View full text Add to dashboard Cite

Motivated by the increasing interest in clinical studies focused on infant movements and poses, this research addresses the limited emphasis on speed and efficiency in existing 2D and 3D pose estimation methods, particularly concerning infant datasets. The scarcity of publicly available infant data poses a significant challenge. In response, we aim to develop a lightweight pose estimation model tailored for edge devices and CPUs. Drawing inspiration from the OpenPose-2016 approach, we refine the algorithm’s architecture, focusing on 2D image training. The resulting model, with 4.09 million parameters, features a single-branch structure. During execution, it achieves an algorithmic complexity of 8.97 giga floating-point operations per second (GFLOPS), enabling operation at approximately 23 frames per second on a Core i5-10400f processor.Notably, this approach balances compact dimensions with superior performance on our self-collected infant dataset. We anticipate that this pragmatic methodology establishes a robust foundation, addressing the need for speed and efficiency in infant pose estimation and providing favorable conditions for future research in this application.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Methodsmentioning

confidence: 99%

Section: Feature Extractormentioning

confidence: 99%

See 1 more Smart Citation