Category Disentangled Context: Turning Category-irrelevant Features Into Treasures

Keke Tang, Guodong Wei, Jie Zhu, Yuexin Ma, Runnan Chen, Zhaoquan Gu, Wenping Wang

Abstract: Deep neural networks have achieved great success in computer vision, thanks to their ability in extracting category-relevant semantic features. On the contrary, irrelevant features (e.g., background and confusing parts) are usually considered to be harmful. In this paper, we bring a new perspective on the potential benefits brought by irrelevant features: they could act as references to help identify relevant ones. Therefore, (1) we formulate a novel Category Disentangled Context (CDC) and develop an adversarial deep network to encode it; (2) we investigate utilizing the CDC to improve image classification with the attention mechanism as a bridge. Extensive comparisons on four benchmarks with various backbone networks demonstrate that the CDC could bring remarkable improvements consistently, validating the usefulness of irrelevant features.

One-sentence Summary: category-irrelvant features are useful.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Reviewed Version (pdf): https://openreview.net/references/pdf?id=UJh5V8qilB