Unsupervised Visual Representation Learning by Context Prediction

Title: Unsupervised Visual Representation Learning by Context Prediction
Authors: Carl Doersch, Abhinav Gupta, Alexei A. Efros
Published: 19th May 2015 (Tuesday) @ 21:18:17
Link: http://arxiv.org/abs/1505.05192v3

Abstract

This work explores the use of spatial context as a source of free and plentiful supervisory signal for training a rich visual representation. Given only a large, unlabeled image collection, we extract random pairs of patches from each image and train a convolutional neural net to predict the position of the second patch relative to the first. We argue that doing well on this task requires the model to learn to recognize objects and their parts. We demonstrate that the feature representation learned using this within-image context indeed captures visual similarity across images. For example, this representation allows us to perform unsupervised visual discovery of objects like cats, people, and even birds from the Pascal VOC 2011 detection dataset. Furthermore, we show that the learned ConvNet can be used in the R-CNN framework and provides a significant boost over a randomly-initialized ConvNet, resulting in state-of-the-art performance among algorithms which use only Pascal-provided training set annotations.

---

Early self-supervised vision paper providing the context prediction pretext task.

• Idea: Self-supervised approach which relies on predicting the spatial relations between two patches of the same image as a pre-text pre-training task - specifically a central patch and one of 8 patches around it arranged in a grid
• Clever tricks:
  • Authors handle “trivial” solutions - model learns to predict spatial relation by learning to find lines at the shared boundaries of patchs by
    • allowing gaps between
    • jittering the surrounding 8 patches
  • Authors handle chromatic abbrration-based trivial solution by
    • channel dropping - randomly drop 1 out of 3 colour RGB channels on a per patch basis
    • Projecting onto the red-green axis by multiplying all colours (length-3 vectors) by a matrix - I think this causes the information from one of the channels to collapse into this axis or similar
• Surprising results:
  • Figure 4: Astonishing that random initialisation of the network yields a (somewhat) effective feature encoder
  • Authors compare their self-supervised approach which relies on predicting the spatial relations