2020 · Paper
On Translation Invariance in CNNs
Overview
This CVPR 2020 paper challenges the fundamental assumption that convolutional neural networks maintain translation invariance. We demonstrate that CNNs can and will exploit absolute spatial location by learning filters that respond exclusively to particular locations through image boundary effects.
Key Contributions
- Demonstrates that CNNs exploit absolute spatial location despite architectural design
- Identifies image boundary effects as the mechanism for position-sensitive learning
- Proposes solutions to improve translation invariance
- Shows benefits for small datasets across classification, patch matching, and video tasks