Object Recognition

Finding and identifying objects within an image is very challenging not only because of the large variety in appearance of objects, but also due to the staggering number of possible locations an object can occupy. Humans have the remarkable ability to quickly find coherent regions in an image which surely facilitates recognition. Recently, impressive progress has been made in computer vision on finding sets of coherent regions that accurately cover objects. By reducing the set of possible interesting locations and providing better object boundaries, these works facilitate the use of more expensive and advanced computer vision techniques on the regions of an image that matter most, giving rise to substantial improvements in object localisation and semantic segmentation. Our research works attept to generate possible object locations with a data-driven selective search approach (shown in Fig. 1) and improve the generation of classindependent object regions (shown in Fig. 2). 

Fig. 1 The training procedure of our object recognition pipeline. As positive learning examples we use the ground truth. As negatives we use examples that have a 20-50% overlap with the positive examples. We iteratively add hard negatives using a retraining phase

Fig. 2 The proposed framework for generating class-independent object hypotheses. We start with an oversegmentation. Then we greedily group segments together in a hierarchical fashion where we use a pre-trained Random Forest classifier to determine at each stage of the hierarchy which regions should be merged. Afterwards we identify stable blobs and merge all adjacent pairs. Merging stable blobs allows for the discovery of objects which consist of visually dissimilar parts.

Selected papers:

  • Selective Search for Object Recognition
    J.R.R. Uijlings, K.E.A. van de Sande, T. Gevers, A.W.M. Smeulders

    International Journal of Computer Vision (IJCV), doi:10.10007/s11263-013-0620-5, 2013.  [PDF][BibTex]

  • Learning to Group Objects
    Victoria Yanulevskaya, Jasper Uijlings, Nicu Sebe
    IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. [PDF][BibTex]
  • A proto-object-based computational model for visual saliency
    Victoria Yanulevskaya, Jasper Uijlings, Jan-Mark Geusebroek, Nicu Sebe, Arnold Smeulders
    Journal of Vision, 13(13):1-19, 2013. [PDF][BibTex]
  • Sparse Color Interest Points for Image Retrieval and Object Categorization
    Julian Stottinger, Allan Hanbury, Nicu Sebe, Theo Gevers
    IEEE Transactions on Image Processing (TIP), 21(5):2681-2692, 2012. [PDF][BibTex]