Reverse-engineer the Distributional Structure of Infant Egocentric Views for Training Generalizable Image Classifiers

TL;DR

This paper investigates the diverse distribution of infant egocentric views, demonstrating how simulating this distribution can enhance the training of more robust and generalizable image classifiers for both infant and third-person vision tasks.

Contribution

It provides empirical evidence of the diversity in infant egocentric views and shows how simulating this distribution improves classifier generalization.

Findings

Infant egocentric views have more diverse distributions than adults' views.

Simulating infant view distributions benefits training of generalizable classifiers.

The approach improves classifier performance in infant and third-person vision.

Abstract

We analyze egocentric views of attended objects from infants. This paper shows 1) empirical evidence that children's egocentric views have more diverse distributions compared to adults' views, 2) we can computationally simulate the infants' distribution, and 3) the distribution is beneficial for training more generalized image classifiers not only for infant egocentric vision but for third-person computer vision.