Ivan Cisneros
Menu
Academic project

Mini Places Challenge - Image Classification

An ensemble scene classifier combining residual networks, object-recognition features, and a Random Forest for the Mini Places Challenge.

  • Machine learning
  • computer vision
  • Residual CNNs
  • ImageNet-trained object recognition
  • Random Forest
  • data augmentation
Date
2016
Organization
Massachusetts Institute of Technology
Role
Image-classification team member on a four-person team
Links
Report (opens in a new tab)Challenge (archived) (opens in a new tab)
Diagram of three scene-classification networks and an object-recognition network feeding a Random Forest classifier.
Scene and object classifiers feed the final ensemble.

Overview

The Mini Places Challenge asked teams to classify photographs into 100 scene categories using 100,000 training images. Our goal was to improve scene recognition by combining information about the overall setting with information about objects visible in it.

Approach

Our team trained three residual convolutional neural networks for scene classification and averaged their predictions. We added outputs from a separate ImageNet-trained object-recognition network so the system could use object cues alongside scene predictions. A Random Forest combined the scene and object-network outputs into final classifications. To limit overfitting on the smaller challenge dataset, we augmented training images with random rotations, translations, shears, and zooms.

Results

The combined system achieved 43.97% Top-1 error and 16.44% Top-5 error on the Mini Places test set.