Ensemble Projection for Semi-supervised Image Classification

Dengxin Dai and Luc Van Gool

Please see the recent results of Ensemble Projection with the CNN features here if you are interested.


This paper investigates the problem of semi-supervised classification. Unlike previous methods to regularize classifying boundaries with unlabeled data, our method learns a new image representation from all available data (labeled and unlabeled) and performs plain supervised learning with the new feature. In particular, an ensemble of image prototype sets are sampled automatically from the available data, to represent a rich set of visual categories/attributes. Discriminative functions are then learned on these prototype sets, and image are represented by the concatenation of their projected values onto the prototypes (similarities to them) for further classification. Experiments on four standard datasets show three interesting phenomena: (1) our method consistently outperforms previous methods for semi-supervised image classification; (2) our method lets itself combine well with these methods; and (3) our method works well for self-taught image classification where unlabeled data are not coming from the same distribution as labeled ones, but rather from a random collection of images.


Ensemble Projection!

Fig1. The pipeline of Ensemble Projection (EP). EP consists in unsupervised feature learning (left panel) and plain supervised classification (right panel). For feature learning, we sample an ensemble of T diverse prototype sets from all known images and learn discriminative classifiers on them for the projection functions. Images are then projected using these functions to obtain their new representation. For classification, we train plain classifiers on labeled images with the learned features to classify the unlabeled ones.


Ensemble Projection!

Fig2. Semi-supervised classification results (average ap) on the four datasets. The top panel evaluate the performance of our learned features when fed into Logistic Regression and SVMs. The bottom shows its performance when fed into HF [34] and LapSVM [1]. All methods were tested with two feature inputs: the concatenation of GIST, PHOG and LBP, and our learned feature from them (indicated by “+ EP”).

Ensemble Projection!

Fig3. Comparison of our learned features (indicated by EP(.)) to the corresponding original features GIST, PHOG, and LBP. Logistic Regression was used as the classifier with 5 labeled training images per class.


The matlab code of ensemble projection and the code I used to extract all the features are available to use.

The datasets Scene-15, LandUse-21, Texture-25, and Caltech-101 were used in the paper, with three common image features GIST, PHOG, and LBP. The features were L1-normalized and then multiplied by their dimensions (e.g. for GIST, the L1-normalized feature values are then multiplied by 20.). This is to map all feature values to comparable levels.
The data (MAT-File) we used for the paper are: 15_gist, 15_phog, 15_lbp, 21_gist, 21_phog, 21_lbp, 25_gist, 25_phog, 25_lbp, 101_gist, 101_phog, 101_lbp.

D. Dai and L. Van Gool. Ensemble Projection for Semi-supervised Image Classification. ICCV, 2013. PDF BibTex

The Poster presented in ICCV 2013.

This page has been edited by Dengxin Dai