NUDT's Participation in the Robot Vision Challenge of ImageCLEF2014

Abstract

This working note describes the method of the NUDT team for scene classi cation and object recognition in the ImageCLEF 2014 Robot Vision Challenge. The method is composed of two steps: 1. spatial pyramid match (SPM) and a Pyramid of HOG (Histograms of Oriented Gradient) are incorporated to represent an indoor place image. 2. a multiclass SVM (Support Vector Machine) is utilized to classify an image by one-versus-all binary SVMs. Based on the method, our system wins the championship this year.

Keywords

• indoor robot localization
• indoor object recognition
• SPM
• PHOG
• multi-class SVM

Introduction

The proposed image representation incorporates appearance and shape. For appearance, we follow the approach of SPM [ 1 ]. SIFT descriptors [2] of 16 16 pixel patches are computed over a dense regular grid with spacing of 8 pixels. We perform k-means clustering of a random subset of patches from the training set to form a visual codebook of V codes, and the corresponding descriptors are assigned to their closest vocabulary codes using the Euclidean distance. In the pyramid layer, multiple codes from inside each sub-region are pooled together into a histogram. Finally, the histograms from all sub-regions are concatenated together to generate SPM. Local place shape is described by HOG [3] within an image block quantized into B bins. Orientations of edges within a certain angular range are counted, forming each bin in the histogram. Incorporating spatial pooling like SPM, PHOG is formed [4]. The nal image representation is formed by concatenating SPM and PHOG. The total cells at level l in the each pyramid is 4l. The entire image representation is a vector with dimensionality

L L V X 4l + B X 4l. For example, levels up to L = 1, V = 10 visual words and l=0 l=0 B = 10 bins it will be a 100-vector. 3

Multiclass classi cation

We believe that adopting an appropriate multi-class classi er meets the requirement of promoting the recognition performance. The multi-class SVM using one-versus-all role proves its discrimination in many practical applications and is adopted in our system. For each class s, after the responding SVM classi er has been obtained through training, the nal decision function fs(x) for a test sample x has the following form:

m fs(x) = max(X ai yi ker(xi; x) + bs)

s i=1 (1) Where ker(xi; x) is the kernel function, ais; bs are the learned model parameters of each one-versus-all binary SVM, yi is the label of the training sample xi, m is the number of the training samples. Then the ultimate label belongs to the class with the maximal value. In practice, we nd that the one-versus-all SVM with nonlinear kernel is suitable for classifying the scenes and objects. 4

Experiments

In the Robot Vision Challenge this year, only one training sequence is provided with 5000 labeled images. An additional validation sequence with 1500 labelled images is provided, and includes 500 images of a non-previously seen building that presents similar room categories and objects as for the training sequences and 1000 images of the old building in the training sequence. The test sequence contains 3000 unlabeled images. In these three sequence, only visual information is considered irrespective of depth information.

For the task, we utilize the whole images of training sequence for training SVM, and the result is similar to the result of utilizing the same number of images for per class to train SVM. In the application, only grey level cues are used. And a visual codebook of 200 codes is used for appearance, and HOG with range [0; 360] using all orientation is discretized into B = 8 bins for shape. For multi-class classi cation, we utilize LIBSVM [5] with one-versus-all classifying rule. Specially, pyramid match kernel [ 1 ] is adopted. Moreover, we limit the number of levels to L = 2 to prohibit over tting. All experiments are repeated 10 times with all training images. Some example images are shown in Fig. 1, Fig. 2, Fig. 3.

Results on the validation sequence We utilize our method on the validation sequence, and compute the respective classi cation accuracy for each scene category and object class. Detailed results are shown in Table 1 and Table 2. In the ultimate submission for classifying test sequence, we mix the training sequence and the validation sequence into a dataset for training and classify the test sequence. Scene classi cation and object recognition are realized separately. Finally, Our team ranks the rst out of four teams, results are listed in Table 3. # 1 2 4

Group

NUDT

UFMS

3 Baseline Results

AEGEAN

In the paper, we introduce our method for the Robot Vision Challenge in ImageCLEF2014. Our method applies the SPM and PHOG for image representation and multi-class SVM for classi cation, which achieves the best performance among all the participants. However, From the results of the challenge, there are still some scene categories and object classes can not be processed correctly. In addition, obviously, the non-previously seen building in uences the performance. In the future, we will focus on the further improvement on the classi cation accuracy and the generalization of our method. Recognition, 2006 IEEE Computer Society Conference on, volume 2, pages 2169{ 2178, 2006.

2. DavidG. Lowe. Distinctive image features from scale-invariant keypoints. Interna

tional Journal of Computer Vision, 60(2):91{110, 2004.

3. N. Dalal and B. Triggs. Histograms of oriented gradients for human detection.

In Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer Society Conference on, volume 1, pages 886{893 vol. 1, June 2005. 4. A. Bosch, A. Zisserman, and X. Muoz. Image classi cation using random forests and ferns. In Computer Vision, 2007. ICCV 2007. IEEE 11th International Conference on, pages 1{8, Oct 2007.

5. Chih-Chung Chang and Chih-Jen Lin. LIBSVM: A library for support vector ma

chines. ACM Transactions on Intelligent Systems and Technology, 2:27:1{27:27, 2011. Software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm.

6. Barbara Caputo, Henning Muller, Jesus Martinez-Gomez, Mauricio Villegas, Burak

Acar, Novi Patricia, Neda Marvasti, Suzan Uskudarl , Roberto Paredes, Miguel

Cazorla, Ismael Garcia-Varea, and Vicente Morell. ImageCLEF 2014: Overview and

analysis of the results. In CLEF proceedings, Lecture Notes in Computer Science.

Springer Berlin Heidelberg, 2014.

7. Jesus Martinez-Gomez, Miguel Cazorla, Ismael Garcia-Varea, and Vicente Morell.

Overview of the ImageCLEF 2014 Robot Vision Task. In CLEF 2014 Evaluation

Labs and Workshop, Online Working Notes, 2014.

References

1. S. Lazebnik , C. Schmid , and J. Ponce . Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories . In Computer Vision and Pattern