Seeing through bag-of-visual-word glasses: towards understanding quantization effects in feature extraction methods Alexander Freytag∗ , Johannes R¨uhle∗ , Paul Bodesheim∗ , Erik Rodner∗ , Joachim Denzler∗ ∗ Computer Vision Group, Friedrich Schiller University Jena


Summary To answer this question, we present an indepth analysis of the effect of local feature quantization on human recognition performance. Our analysis is based on recovering the visual information by inverting quantized local features and presenting these visualizations with different codebook sizes to human observers (Fig. 2). Although feature inversion techniques are around for quite a while, to the best of our knowledge, our technique is the first visualizing especially the effect of feature quantization. Thereby, we are now able to compare single steps in common image classification pipelines to human counterparts. Our results show that (i) humans perform significantly worse than machine learning approaches when being restricted to the visual information present in quantized local features rather than having access to the original input images, and (ii) that early stages of low level local feature extraction seem to be most crucial with respect to achieving human performance on original images. Finally, we demonstrate (iii) that large codebook sizes in the order of thousands of prototypes are essential not only for good machine learning performance, but more interestingly, also for human image understanding. Method Our technique is simple and in line with current trends in image reconstruction from local features. For an unseen image, we extract local features on a dense grid and follow the bag-of-words paradigm by quantizing them using a pre-computed codebook. Based on inversion techniques for local features, we can compute the most probable image patch for every prototype, i.e., we can visually inspect the quantization quality for a given codebook. Thus, for any local feature, we vector-quantize it with a codebook and draw the inverted prototype into the reconstruction

Local descriptors

Vector quantization















Index of Prototype




su O al ur iz at io n





average recognition rate [%]

Input image


BoW feature

Problem formulation Vector-quantized local features frequently used in bag-of-visual-words approaches are the backbone of popular visual recognition systems due to both their simplicity and their performance. Despite their success, standard bag-ofwords-histograms basically contain low-level image statistics (e.g., number of edges of different orientations). The question remains how much visual information is lost in quantization when mapping visual features to visual “words” and elements of a codebook?

Spatial pooling

80 70

ML-performance (k=2048) ML-performance (k=32)

60 50 40 30 20 10 0

Random guessing 32






type of quantization / codebook size

Fig. 1: Inversion strategy and human classification result in comparison to machine performance

Fig. 2: Overview of the images presented to human observers during the experiment (15 Scenes dataset) image with position and size according to the support of the extracted local feature (Fig. 1). In contrast to previous works for feature inversion, we aim at explicitly visualizing quantization effects. For the simplicity of demonstration, we use HOG features, where code for the inversion technique is publicly available1 . Evaluation We performed human studies, where more than 3,000 images were classified by several individuals with different levels of quantization (Fig. 2). The results are given in the right plot of Fig. 1 and will be presented in detail at the poster2 . 1 2 Source code, detailed results, and a web interface to our evaluation server are available at representation .

towards understanding quantization effects in feature ...

Seeing through bag-of-visual-word glasses: towards understanding quantization effects in feature extraction methods. Alexander Freytag∗, Johannes Rühle∗, ...

2MB Sizes 1 Downloads 59 Views

Recommend Documents

Towards an understanding of oversubscription in cloud
Cloud providers oversubscribe their data centers to lever- ... We consider an Infrastructure as a Service (IaaS) cloud ..... Then, the total utility of the system is.

mail, making telephone call, reading papers, writing articles, and so on. Generally ...... Reunolds, P. [2003] Call Center Staffing (The Call Center School Press,.

Human Factors of Automated Driving: Towards Predicting the Effects ...
Interface. Human driving behaviour. Longitudinal and lateral dynamics. Vehicle. Road and traffic flow conditions. Driver capabilities. Environmental conditions ... Congestion. Automated driving. What are the effects on traffic flow efficiency? Accide

fractional quantization
One of my favorite times in the academic year occurs in early spring when I give my .... domain walls between even-contracted regions and odd-contracted ones. ..... a shift register, the net result being to transfer one state per Landau level.

Towards Understanding Software Evolution: One-Line ...
[email protected] Dewayne E. Perry. ECE & UT ARISE. The University of Texas at Austin. Austin, Texas 78712-1084 [email protected] Abstract. Understanding the impact of software changes has been a challenge since software systems were first

Features in Concert: Discriminative Feature Selection meets ...
... classifiers (shown as sample images. 1. arXiv:1411.7714v1 [cs.CV] 27 Nov 2014 ...... ImageNet: A large-scale hierarchical im- age database. In CVPR, 2009. 5.

Introducing Temporal Asymmetries in Feature ...
improvement in phoneme error rate on TIMIT database over the MRASTA technique. Index Terms: feature extraction, auditory neurons, speech recognition. 1.