Can Simulated Retinal Wave pre-training improve Optical Flow Estimation? - Inria - Institut national de recherche en sciences et technologies du numérique
Poster Année : 2024

Can Simulated Retinal Wave pre-training improve Optical Flow Estimation?

Résumé

The objective of this work is to develop a machine learning model for Optical Flow Estimation (OFE) that employs simulated stage II Retinal Waves as a training set. OFE is a crucial task in computer vision used to infer the motion between two consecutive frames in a video. Training data-driven OFE models is challenging, and the difficulty of obtaining dense Optical Flow ground-truth from real-world video recordings further complicates this task. Consequently, all state-of-the-art OFE models mostly train on synthetic data, which often introduces biases. We ask the question: Could simulated Retinal Waves improve the performance or efficiency of such data-driven OFE models? Retinal Waves are one of the early processes in visual system development, occurring around birth and playing a vital role in the structuring of retinal circuits as well as their projections to the visual pathway. These waves are spontaneous bursts of action potentials in the immature retina, forming activity waves in abstract patterns, and prepare the visual system to detect motion before the eyes open. We adopt a biophysical model [1] to generate biologically-plausible waves, which we use to pre-train the encoder of a recurrent architecture that is subsequently fine-tuned for varied OFE benchmarks. The poster will present the developed machine learning workflow and some preliminary results, highlighting the potential benefits of using simulated Retinal Waves in training OFE models. [1] B. Cessac and D. Matzakos-Karvouniari. “The non linear dynamics of retinal waves ”. In: Physica D: Nonlinear Phenomena (2022). doi.org/10.48550/arXiv.2308.12579.
Fichier principal
Vignette du fichier
poster_CK.pdf (1.28 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-04735162 , version 1 (14-10-2024)

Licence

Identifiants

  • HAL Id : hal-04735162 , version 1

Citer

Christos Kyriazis, Bruno Cessac, Hui-Yin Wu, Pierre Kornprobst. Can Simulated Retinal Wave pre-training improve Optical Flow Estimation?. Neuromod meeting 2024, Jul 2024, Antibes, France. ⟨hal-04735162⟩
0 Consultations
0 Téléchargements

Partager

More