| CALVIN group Objectness measure V1.5
Bogdan Alexe, Thomas Deselaers, Vittorio Ferrari
Overview
What is objectness?
The objectness measure acts as a class-generic object detector. It quantifies how likely it is for an image window to contain an object of any class, such as cars and dogs, as opposed to backgrounds, such as grass and water. We release here software for computing objectness [1,2] and sampling any desired number of windows from an image according to their probability of containing an object.
For applications, we recommend to sample about 1000 windows, which ensures covering most objects even in very difficult images (e.g. with small objects and lots of clutter). However, in images of normal difficulty 100 windows are sufficient (e.g. images downloaded from image search engines).
Version V 1.5 includes a new window sampling strategy (NMS) which leads to higher detection rates. On the highly challenging PASCAL VOC 2007 dataset [3], the top 1000 sampled windows now cover 91% of all objects [2], as opposed to about 70% in previous versions using multinomial sampling [1].
In addition to the source code, we also release sampled windows for every image from PASCAL VOC 2007 [3], for both the new NMS sampling strategy and for the older multinomial sampling. These ready-to-use windows hopefully will facilitate applications on this dataset.
How fast is it?
Objectness is computationally efficient. On a mid-range PC, it takes 4 seconds to compute the objectness measure and to sample 1000 windows, for an image of size 350 x 500.
Applications of objectness
Objectness is intended as a low-level preprocessing stage, to propose a small number of windows likely to cover all objects in the image. It has been used in several applications so far:+ learning object classes from weakly supervised images [4,5,12]
+ weakly supervised segmentation of a single object class [6] and of multiple classes [7]
+ unsupervised object discovery [8]
+ learning spatial relations between objects and humans in human-object interactions [9]
+ content-aware image resizing [10,11]
+ speeding up class-specific detectors [1,2]
+ reducing the false-positive rates of class-specific detectors [1,2]
+ object tracking in video (ongoing work)
+ building block for other class generic detectors [13]
Examples
For each of image we show the windows best covering the objects annotated in the PASCAL VOC 2007 (our of 1000 windows sampled with NMS). We mark in yellow windows correctly covering ground-truth objects (cyan); if there is more than one correct window, the best one is shown.
Downloads
| Filename | Description | Size |
|---|---|---|
| Source code (Matlab/C) | Souce code for objectness measure | 20 MB |
| README.txt | Description of content | 9 kB |
| LICENSE | Software license | 1 kB |
| PASCAL VOC 2007 windows using NMS sampling | 1000 windows sampled using the NMS strategy for each image in PASCAL VOC 2007 (recommended) | 86 MB |
| PASCAL VOC 2007 windows using multinomial sampling | 10000 windows sampled using the multinomial strategy for each image in PASCAL VOC 2007 | 576 MB |
New in V 1.5
+ NMS sampling procedure added
New in V 1.01
+ windows with width or height = 1 pixel are not anymore considered
Publications
[1] Alexe, B., Deselares, T. and Ferrari, V.
What is an object?
CVPR 2010.
Document: PDF
[2] Alexe, B., Deselares, T. and Ferrari, V.
Measuring the objectness of image windows
PAMI 2012.
Document: PDF
[3] Everingham, M., Van Gool, L., Williams, C., Winn, J., and Zisermann, A.
The PASCAL Visual Object Classes Challenge 2007.
[4] Deselares, T., Alexe, B. and Ferrari, V.
Localizing objects while learning thier appearance
ECCV 2010.
[5] Khan, I., Roth, P. M. and Bischof, H.
Learning Object Detectors from Weakly-Labeled Internet Images
OAGM Workshop 2011.
[6] Alexe, B., Deselaers, T. and Ferrari, V.
ClassCut for unsupervised class segmentation
ECCV 2010.
[7] Vezhnevets, A., Ferrari, V. and Buhmann, J.
Weakly supervised semantic segmentation with a multi-image model
ICCV 2011.
[8] Lee, Y. J. and Grauman, K.
Learning the easy things first: Self-paced visual category discovery
CVPR 2011.
[9] Prest, A., Schmid, C. and Ferrari, V.
Weakly supervised learning of interactions between humans and objects
PAMI 2011.
[10] Sun, J., and Ling, H.
Scale and Object Aware Image Retargeting for Thumbnail Browsing
ICCV 2011.
[11] Bao, X., Narayan, T., Sani, A. A., Richter, W., Choudhury, R. R., Zhong, L. and Satyanarayanan, M.
The Case for Context-Aware Compression
ACM Hotmobile 2011.
[12] Siva, P. and Xiang, T.
Weakly Supervised Object Detector Learning with Model Drift Detection
ICCV 2011.
[13] Rahtu, E., Kannala, J. and Blaschko, M.
Learning a Category Independent Object Detection Cascade
ICCV 2011.
Acknowledgements
This work is funded by the Swiss National Science Foundation SNSF