Original title: The Blind Spot in 2D Infants' Pose Estimation:Robust Learning from Noisy Annotations
Emanuele Cardinale, Marco Proietti, Alessandro Cacciatore, Maria Francesca Spadea +2 · September 2026
Tracking how preterm babies move helps doctors evaluate their neurological development. To do this, computer vision models track key points on the baby's body, like joints. However, training these models requires accurately labeled images. In hospitals, caregivers block the view or babies hide their own limbs, leading to human labeling mistakes that confuse the models. To fix this problem, researchers created a method called REMIND. REMIND tracks how the model's predictions for each body point change over time during training. By analyzing these learning patterns, it groups together and spots the wrongly labeled points without needing to know what kinds of errors were made beforehand. It then removes those faulty labels so the model learns cleanly. Tested on real hospital videos of preterm infants, REMIND successfully identified inaccurate annotationsLabels marking keypoints on images, such as where a baby's limbs are. across several simulated error conditions, helping build reliable tools for infant health monitoring even when data has mistakes.
Imagine you are trying to draw a map of a busy playground, but kids keep moving, so some marks on your map are wrong. The hard part is figuring out which marks are accurate without checking each one by hand. In this paper, the method looks at how the drawing process changes over time to guess which marks are likely mistakes. It helps clean up the map so you can use it reliably.
Doctors monitoring preterm infants rely on accurate movement data to assess neurodevelopment, but labeling mistakes can corrupt this data. This method helps ensure the data is reliable, supporting better clinical decisions.
The abstract does not explicitly state limitations, but evaluation was on a proprietary dataset (NeoPose), so generalizability to other settings is not established. It is also noted as the first study to address this problem, implying further validation may be needed.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
The Blind Spot in 2D Infants' Pose Estimation:Robust Learning from Noisy Annotations — Emanuele Cardinale, Marco Proietti, Alessandro Cacciatore et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.