A Random CNN Sees Objects: One Inductive Bias of CNN and Its Applications

TL;DR

This paper reveals that randomly initialized CNNs inherently focus on objects, a bias called Tobias, which can be exploited to improve self-supervised learning and object detection tasks.

Contribution

The paper introduces Tobias, an inductive bias in CNNs that naturally localizes objects without training, and demonstrates its application to enhance self-supervised learning and detection performance.

Findings

Random CNNs can localize objects without training.

Tobias improves downstream object detection tasks.

Tobias is robust across different training set sizes and augmentations.

Abstract

This paper starts by revealing a surprising finding: without any learning, a randomly initialized CNN can localize objects surprisingly well. That is, a CNN has an inductive bias to naturally focus on objects, named as Tobias ("The object is at sight") in this paper. This empirical inductive bias is further analyzed and successfully applied to self-supervised learning (SSL). A CNN is encouraged to learn representations that focus on the foreground object, by transforming every image into various versions with different backgrounds, where the foreground and background separation is guided by Tobias. Experimental results show that the proposed Tobias significantly improves downstream tasks, especially for object detection. This paper also shows that Tobias has consistent improvements on training sets of different sizes, and is more resilient to changes in image augmentations. Code is available at https://github.com/CupidJay/Tobias.