Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Viewpoint Diversity Improves Convolutional Neural Network Generalization and Robustness

Yifan Luo¹, Niklas Müller¹; ¹University of Amsterdam

Presenter: Yifan Luo

Although convolutional neural networks (CNNs) reach human-level accuracy on standard object recognition tasks, they perform poorly when faced with changes in viewpoint or corrupted images. In this study, we demonstrate that these two distinct failure modes can be addressed using a single strategy: training on diverse viewpoints. To investigate this, we created artificial image datasets that systematically vary in viewpoint diversity while keeping the dataset size constant, to train and evaluated CNN object recognition performance. Our results reveal a core trade-off between learning speed and generalization performance. On the one hand, models trained on restricted viewpoints exhibit fast learning and achieve near-perfect in-distribution accuracy, but they overfit to specific views, resulting in dramatic performance drops on unfamiliar viewpoints. On the other hand, training with diverse viewpoints slows learning but significantly improves out-of-distribution performance. Notably, exposure to diverse viewpoints also greatly enhances robustness to common image corruptions. These results point to a shared mechanism for achieving robustness to both viewpoint variation and image corruption, and further alignment with human performance.

Topic Area: Visual Processing & Computational Vision

Extended Abstract: Full Text PDF