Presentation is loading. Please wait.

Presentation is loading. Please wait.

A. M. R. R. Bandara & L. Ranathunga

Similar presentations


Presentation on theme: "A. M. R. R. Bandara & L. Ranathunga"— Presentation transcript:

1 A. M. R. R. Bandara & L. Ranathunga
Feature Clustering Approach Based on Histogram of Oriented Flow and Superpixels A. M. R. R. Bandara & L. Ranathunga Department of Information Technology, Faculty of Information Technology, University of Moratuwa, Sri Lanka. N.A. Abdullah Department of Computer System and Technology, Faculty of Computer Science & Information Technology, University of Malaya, Malaysia. Financial Support by National Research Council Sri Lanka Grant No

2 Overview Introduction Related Studies Proposed Method
Experimental Results & Discussion Conclusion

3 Introduction Object Localization Exhaustively search over video frames
Repetitively constructs visual feature descriptors Classify each of the descriptors until an object is found. Problems: Feature extraction and description are computationally expensive Exhaustive search prevents real time processing Supervised methods require to re-supervise when the application domain is changed. Object segmentation Rely on several assumptions Static camera Objects are either static or moving Number of objects in the scene is known An objects contain uniformly distributed colors

4 Introduction Objectives of the Study
Evaluate the suitability of a new feature namely Salient Dither Pattern Feature (SDPF) for segmenting semantically meaningful objects. Efficiently segment objects by clustering SDPF feature points in an environment which is having: Heavy camera movements Both static and moving objects

5 Related Studies The study depends on three main previous studies
Salient Dither Pattern Feature (SDPF) Simple Linear Iterative Clustering (SLIC) Superpixel Histogram of Oriented Flow (HOOF)

6 Salient Dither Pattern Feature (SDPF)
A previous work called Compacted Dither Pattern Code (CDPC) has proved that dithering can reduce visual data without affecting the overall visual impression[1]. SDPF utilizes the same concept to derive a point like feature. Previous studies has demonstrated that SDPF is highly robust to rotational and scale transformation More details can be found in: A. Bandara, L. Ranathunga, and N. Abdullah, “Invariant properties of a locally salient dither pattern with a spatial-chromatic histogram,” in Industrial and Information Systems (ICIIS), th IEEE International Conference on, 2013, pp. 304–308. [1]. L. Ranathunga, R. Zainuddin, and N. A. Abdullah, “Performance evaluation of the combination of Compacted Dither Pattern Codes with Bhattacharyya classifier in video visual concept depiction,” Multimed. Tools Appl., vol. 54, no. 2, pp. 263–289, 2011.

7 Simple Linear Iterative Clustering (SLIC) Superpixel
Superpixels capture redundancy in an image and greatly reduce the complexity of subsequent image processing tasks. Obtained by performing a local clustering (K-Means) of pixels in the 5-D space: L, a, b in CIELAB and x,y pixel coordinates. SLIC shows a strong perceptual homogeneity within the superpixels and O(n) time complexity Have proved increasingly useful for semantic segmentation [1]. [1]. A. Papazoglou and V. Ferrari, “Fast object segmentation in unconstrained video,” in Computer Vision (ICCV), 2013 IEEE International Conference on, 2013, pp. 1777–1784.

8 Histogram of Oriented Flow (HOOF)
A time series of HOOF can characterize a high-level object motion [1] Can be built by binning each flow vector according to its primary flow angle and weighted according to its magnitude. HOOF feature is robust to the flow noises. [1]. R. Chaudhry, A. Ravichandran, G. Hager, and R. Vidal, “Histograms of oriented optical flow and binet-cauchy kernels on nonlinear dynamical systems for the recognition of human actions,” in Computer Vision and Pattern Recognition, CVPR IEEE Conference on, 2009, pp. 1932–1939.

9 Proposed Method Extract SDPF Calculate optical flow
Convert to CIE Lab color space Extract SDPF Calculate optical flow Group the SDPF points to their superpixels Obtain superpixels by SLIC Apply color weighting L’ab Repeat for a given depth of history Tf Obtain HOOF for each group of SDPF points Calculate mean color for each superpixel Obtain the time series of HOOFs for each group Merge the groups with similar mean color of superpixels or similar time series of HOOF To the construction of descriptors for each group of SDPF

10 Experimental Results and Discussion
Method Average Completeness Average Spatial Accuracy Error RSHC 87% 27% K-Means-8D 86% 31% EM-8D 81% 43%

11 Experimental Results and Discussion
Original Ground-truth K-Means-8D RSHC

12 Conclusion The study has proved that the proposed scheme works better than K-Means-8D and EM-8D clustering. The new approach does not required to input the number of objects. The performance evident that the proposed scheme can be used to cluster the SDPF feature points to semantically meaningful visual segments without supervising the clustering method. Limitations: Smaller objects tend to ignore in clustering

13 Questions? Q/A Thank You.


Download ppt "A. M. R. R. Bandara & L. Ranathunga"

Similar presentations


Ads by Google