Skip to main content Skip to main navigation menu Skip to site footer

Location Estimation of a Photo: A Geo-signature MapReduce Workflow

  • Pakpoom Mookdarsanit King Mongkut’s University of Technology North Bangkok
  • Montean Rattanasiriwongwut King Mongkut’s University of Technology North Bangkok

Downloads

Download data is not yet available.

Abstract

Location estimation of a photo is the method to find the location where the photo was taken that is a new branch of image retrieval. Since a large number of photos are shared on the social multimedia. Some photos are without geo-tagging which can be estimated their location with the help of million geo-tagged photos from the social multimedia. Recent researches about the location estimation of a photo are available. However, most of them are neglectful to define the uniqueness of one place that is able to be totally distinguished from other places. In this paper, we design a workflow named G-sigMR (Geo-signature MapReduce) for the improvement of recognition performance. Our workflow generates the uniqueness of a location named Geo-signature which is summarized from the visual synonyms with the MapReduce structure for indexing to the large-scale dataset. In light of the validity for image retrieval, our G-sigMR was quantitatively evaluated using the standard benchmark specific for location estimation; to compare with other well-known approaches (IM2GPS, SC, CS, MSER, VSA and VCG) in term of average recognition rate. From the results, G-sigMR outperformed previous approaches.

1. Introduction

Nowadays, influence of social multimedia has rapidly changed the world. Social multimedia is a virtual world that enables users to share their articles, photos, videos, etc. Social users have built a large number of information stored at the datacenter. IDC Digital Universe forecasted the amount of information over the social multimedia will be 40ZB within 2020 [ 1 ]. Since there are a large number of information, especially in type of photography that are easily taken from the mobile devices; and quickly shared over the social multimedia during their travels. Million photos on the social multimedia are taggedwith their geo-locations (also called geo-tagged photos) which are useful for image retrieval [ 2 ] in terms of browsing, searching, mining, and organizing. However, there are many photos without geo-tagging(also called geo-untagged photos) that are shared on the social multimedia. The EXIF data is automatically tagged during the capturing a photo for description of the location. However, the photos with enhancement process (such as cropping, blending, sharpening and other effects) are easy for loss of EXIF data. To that end, it is feasible that some geo-untagged photos can be estimated their locations with the help of another million geo-taggedphotos from the large-scale dataset. The methodology for finding the geographical location of a photo (where it was taken) is called “location estimation of a photo” which is a new branch of image retrieval. Visual content (such as color, texture, and shapes) and other textual metadata (such as annotations, tags, duplicated comments and/or previous user’s sharing) of a geo-untagged photo can be used to retrieve any similar scenes from the dataset. The location of geo-untagged photo is estimated by the location of the most similar geo-tagged photo (or the most similar group that is categorized from geo-tagged photos) from the dataset. In 2008, the first groundwork was found by Hays and Efros [ 3 ]. They developed a simple image retrieval for automatic geo-tagging of their photos (well-known as IM2GPS) that they were taken with the quote-worthy question as “What can you say about where these photos were taken?”

Later, most researches about the location estimation of a photo still focused on the recognition performance by a traditional spatial coding (named SC) [ 4 ] and an indexing of hierarchical global feature clustering with local feature refinement under the measurement of cosine-based similarity (named CS) [ 5 ]. Some approaches adapt the visual synonym as in the maximally stable extremal region (MSER) algorithm for the salient region mining [ 6 ]. Recently, the visual spatial contents arrangement (named VSA) [ 7 ] and the group of visual contents (named VCG) [ 8 ] from a photo were built to matched the similar scenes from the dataset in 201a4 and 2015, respectively.

Fig 1 In case of the error from similarity-based location estimation of a photo

However, those previous approaches have a common ignorance. Since their location estimations are computed from a geo-tagged photo or a group of geo-tagged photos from the dataset that have the highest number of similar visual contents with the geo-untagged photo. This idea is found that it is easy to produce the error because the most similarity sometimes cannot directly estimate the target location of a photo.

As in St. Petersburg, Russia) and B (Venice, Italy). Since the geo-untagged photo was taken at Venice as the same location as the scene B. Unfortunately, the scene A has a more number of similar contents than the scene B. Hence, location of the geo-untagged photo is wrongly estimated as taken from St. Petersburg as the same location as scene A. Fig. 1 , the geo-untagged photo is used to find the similar scenes from the dataset which are retrieved as the result of scene A (

Location estimation of a photo needs a more procedure to define the uniqueness of a place that to be distinguished from other places. The contribution of our paper is to design a workflow named G-sigMR (Geo-signature MapReduce) for the improvement of recognition performance. First, the important visual contents are selected from a geo-untagged photo (in term of a vector). Second, the system applies the Map-Reduce indexing to map only the similar scenes from the millions of geo-tagged photos in the large-scale dataset; and store them in the candidate space. In contrast, some irrelevant scenes are reduced. Third, all scenes in the candidate space are grouped together. The scenes taken from the same location are grouped into the same group. Fourth, each group is defined the uniqueness of the group named Geo-signature which is summarized from their visual synonyms. Finally, location of the geo-untagged photo is estimated by the group that has the maximal signature. Our G-sigMR is compared to the previous approaches [ 3 , 4 , 5 , 6 , 7 , 8 ] based on Geo-tagged large-scale dataset (GOLD) [ 9 , 10 ] which contains 3.3 millions of geo-tagged photos. From the results, our G-sigMR had the average recognition rate more than 90% which outperformed other approaches.

The remainder of this paper is organized as follows. Section 2 describes crucially about location estimation of a photo and its paradigms. The architecture of our G-sigMR is step-by-step explained in the section 3. The experiments and comparisons are in the section 4. For the conclusion, we summarize our G-sigMR workflow with its modern applications and the direction of our future work in the section 5.

2. Location Estimation of a Photo

Location estimation of a photo is the method to find the geographical location where the photo was taken. Since many million photos are shared by users on the social multimedia. Many photos are taken frequently from the same place. Some of them are either with or without geo-tagging. It is feasible that some geo-untagged photos can be estimated their locations with the help of million geo-tagged photos from the large-scale dataset. The researches about location estimation of a photo can be categorized into 2 paradigms: Visual-based and multisource-based evidence, respectively.

2.1 Visual-based Evidence

The photo representation (in term of visual contents) is considered to find the location. The first process known as feature detection is used to find the photo’s characteristics (also known as key-points) that enable the photo to be distinguished from other photos [ 11 ]. The key-points of the same architectural scene or geography are not changed despite of the diversity of camera-viewpoints. However, feature detection produces only horizontal and vertical co-ordinates of the key-points from a photo [ 12 ] which are not enough for the landmark retrieval. It needs a descriptive vector (known as feature description) to describe the region surrounding the key-point [ 11 , 12 , 13 ] (also known as spatial information [ 14 , 15 , 16 , 17 ]). Feature descriptions of a photo are in term of a vector that represents the important visual contents of the architectural place or geography from a single photo [ 18 , 19 , 20 ]. Geo-tagged photos (in terms of vectors) and their locations are collected in the dataset [ 18 , 19 ] which later can be used in the landmark retrieval system for the indexing of many scenes with their locations [ 21 , 22 ]. For estimating the location of a geo-untagged photo, the most similar scene from the dataset is estimated that it was taken at the same location as the geo-untagged photo [ 23 ]. The most feature descriptions for the computation of vector similarity in the landmark retrieval are SIFT (Scale-invariant Feature Transform) [ 24 , 25 , 26 ], SURF (Speeded-Up Robust Features) [ 27 ] and HoG (Histogram of Gradients) [ 28 , 29 ]. However, those traditional feature descriptions are still having some deficiency in a vector representation of an architectural place or geography for landmark retrieval such as quantization loss, non-discrimination of vector [ 9 , 14 , 15 ]. Most improved methods used the concept of visual synonym to be accomplished geometric coherence estimation [ 30 , 31 , 32 , 33 , 34 ]. A visual synonym is a pair of visual contents [ 35 , 36 , 37 , 38 ] that can be combined to find the location of a photo from the large-scale dataset by mapping to the similar scenes and reducing the irrelevance.

2.2 Multisource-based Evidence

Location of a photo can be estimated using visual contents of the photo in combination with textual metadata which is called Multisource-based evidence. Textual metadata consists of annotations [ 39 , 40 , 41 ], tags [ 42 , 43 ], duplicated comments [ 44 , 45 , 46 , 47 ] and/or previous user’s sharing [ 48 ] that are crawled from the social multimedia. The visual contents of a photo with its textual metadata are used to estimate the location of the photo [ 49 , 50 ]. For example, there is a coal-fire photo with many frequent comments about “Gate Hell” or “Door to Hell”, the location (that the photo was taken) may be at Derweze, Turkmenistan. In contrast, multisource-based evidence exists totally deficiency. Since textual metadata are manually organized by million sharing on the social multimedia which affect to the correctness of textual information. If the frequent texts are not associated with the photo, it is easier to be higher errors in location estimation of a photo. Multisource-based evidence must have a thorough filtering which takes hugely more run-time than visual-based evidence. However, location estimation with the help of textual metadata will be work if the sufficiency of information on the social multimedia.

3. A Geo-signature MapReduce Workflow

Fig 2 The architecture of the G-sigMR workflow

This section technically describes the architecture of Geo-signature MapReduce (G-sigMR) workflow which is a visual-based evidence as shown in Fig. 2. The G-sigMR can be organized into 5 main procedures: input a photo, photo division, vector representation, MapReduce indexing and geo-signature generation, respectively.

3.1 Input a Photo

Either geo-tagged or geo-untagged photo can be input to the workflow. In case of a geo-tagged photo, the photo with its geo-tagging will be incrementally added to the set of large-scale samples. In contrast, a geo-untagged photo will be estimated its location.The photo can be taken from any camera-viewpoint. Objects within the photo also can be located in any position as shown in Fig. 3.

Fig 3 A geo-tagged or untagged photo input to the workflow

3.2 Photo Division

The geo-tagged (or untagged) photo is divided into k parts. Gradients of the photo in the x-axis (Gx) and y-axis (Gy) are computed. Since the gradients are intensity of parts through the photo (also called gradient orientations) which are determined from the high-contrast visual contents such as edges or corners of architectural-objects within the photo. Range of gradients can be mathematically formulated according to Gaussian distribution (-1, 0, 1) in both horizon (denoted as Gx=[-1 0 1]) and vertical (denoted as a transpose of horizon Gy=[-1 0 1]T). And the photo is divided into k parts as shown in Fig. 4. , where k=4 .

Fig 4 A photo division into k parts ( k=4)

After the photo division into k parts, the magnitude of gradient in each part ( G x,y (i) ) is recursively computed by Eq. (1).

where i is a sequence of parts ( 1 ≤ i ≤ k ), k is a number of parts within the photo, G x at i is a magnitude of gradient in the x-axis of the i-th part, G y at i is a magnitude of gradient in the y-axis of the i-th part and G x,y ( i+1) is a magnitude of gradient of the next part (the (i+1)-th part).

The gradient orientation in each part ( x,y (i) ) is also computed in term of the slope of x and y axis by Eq. (2).

where i is a sequence of parts ( 1 ≤ i ≤ k ), k is a number of parts within the photo, G x at i is a magnitude of gradient in the x-axis of the i-th part, G y at i is a magnitude of gradient in the y-axis of the i-th part and x,y (i+1) is an orientation of gradient of the next part (the ( i+1)-th part).

3.3 Vector Representation

Since the important (also called high-contrast) visual contents of a photo need a mathematical representation in term of a vector. From k sub-parts ( k =4). In other words, the photo is further divided into k 2 sub-parts. Fig. 5. , each part from the previous procedure is divided into

Fig 5

Each part division into k sub-parts (known as k2 sub-parts of the photo)

The intensity from pixels in each sub-part within a part is also recursively computed using the overlapping of local contrast normalization with the measure of intensity between parts to be robust for changeable in terms of illumination, scale, rotation and shadowing,as in a series of the orientation of gradient in each sub-part ( q x,y (i,j) ) is recursively computed by Eq. (3).

where i is a sequence of parts ( 1 ≤ i ≤ k ), k is a number of parts within the photo, j is a sequence of sub-parts ( 1 ≤ j ≤ k 2) within the i-th part, k 2 is a number of sub-parts within the photo, B is a set of important visual contents, C x at j is a gradient of the j-th sub-part in the x-axis, C y at j is a gradient of the j-th sub-part in the y-axis, q x,y (i,j+1) is a series of the direction of gradient in the i-th part and the next sub-part (the (j +1)-th sub-part) and q x,y (i+1,j) is a series of the orientation of gradient in next the part (the ( i+1)-th part) and the j-th sub-part.

Only the important (or high-contrast) visual contents from the photo are represented in term of a vector as computed by Eq. (4). that considers from the magnitude of gradient in the i-th part ( G x,y (i) , from Eq. (1). ), the orientationof gradient in the i-th part ( x,y (i) , from Eq. (2). ) and the series of orientation of gradient in the j-th sub-part ( q x,y (i,j) , from Eq. (3). ), respectively.

where i is a sequence of parts ( 1 ≤ i ≤ k ), j is a sequence of sub-parts ( 1 ≤ j ≤ k 2) within the i-th part, k 2 is a number of sub-parts within the photo and represent(j+1) is a representation of important visual contents from the next sub-part (the (j+1)-th sub-part).

3.4 MapReduce Indexing

Customarily, the purpose of “MapReduce approach” [ 51 , 52 , 53 ] was designed for the actual “dynamic structure” for the “velocity”, “volume” and “variety” of non-volatile large-scale datasets (or Big data [ 1 ]). In case of a geo-untagged photo, MapReduce indexing filters only the useful geo-tagged photos (that were collected in term of vectors with geo-tagging) from the set of large-scale samples which are similar to some visual contents of geo-untagged photo. The similarity between the geo-untagged photo and some geo-tagged photos from the set of large-scale samples is recursively computed using index(P k , P * , a) as Eq. (5).

where k is a sequence of geo-tagged photos from the dataset( 1 ≤ k ≤ a ), a is a number of geo-taggedphotos in the set of large-scale samples, P * is the geo-untagged photo, P k is the k-th geo-tagged photo from the set of large-scale samples and index(P k+1 , P * , a) is a similarity between the next geo-tagged photo (the (k+1)-th geo-tagged photo) from the set of large-scale samples and the geo-untagged photo.

For the one-by-one MapReduce structure, the function named map(F i , F j , m, n) is used to match the similar contents between the geo-untagged photo and the k-th geo-tagged photo from the set of large-scale samples. Some geo-tagged photos that have similar visual contents with the geo-untagged photo are chosen to store in the candidate space. If a content F i (from the geo-untagged photo) is not similar to a content F j (from the k-th geo-tagged photo), the reduce(F i , F j ) function eliminates them before execution in the geo-signature computation (the next procedure). In case of unknown (or unseen) location of a geo-untagged photo, the system generates the label of this photo with its visual contents (instead of Geo-signature generation) which will be added to the set of large-scale samples for the next time MapReduce Indexing of this similar photo.

The one-by-one MapReduce structure can be mathematically described in Eq. (6).

where i is a sequence of the i-th visual content of the geo-untagged photo, j is a sequence of the j-th visual content of the k-th geo-tagged photo from the set of large-scale samples, m is a number of visual contents within the geo-untagged photo, n is a number of visual contents within the k-th geo-tagged photo from the set of large-scale samples, F i is the i-th visual content of the geo-untagged photo, F j is the j-th visual content of the k-th geo-tagged photo from the set of large-scale samples, F s is the visual content similarity between the geo-untagged photo and the k-th geo-tagged photo from the set of large-scale samples, n(F s ) is a number of F s (s) , map(F i+1 , F j , m, n) is matching between the next visual content (the (i+1) –th visual content) of the geo-untagged photo and the j-th visual content of the k-th geo-tagged photo from the set of large-scale samples and map(F i , F j+1 , m, n) is matching between the i-th visual content of the geo-untagged photo and the next content (the (j+1) –th visual content) of the k-th geo-tagged photo from the set of large-scale samples.

3.5 Geo-signature Generation

Some geo-tagged photos from the set of large-scale samples have been chosen (from the previous MapReduce indexing) to store in the candidate space because their visual contents are similar with some contents of geo-untagged photo. Within the candidate space, this procedure groups the scenes that were taken from the same location into the same group as shown in Fig. 6.

Fig 6

Grouping of the geo-tagged photos from the candidate space that taken from the same location

The uniqueness (known as unique evidences) of the i-th group named Geo-signature ( GeoSig i) is summarized from photos within the i-th group in term of visual synonyms that are associated with the important visual contents of geo-untagged photo. In other words, the Geo-signature of each group is generated from the similarity of geo-untagged photo which can be generated from Eq. (7).

where i is a sequence of the i-th group, j is a sequence of the j-th geo-tagged photo within the i-th group, m is a number of geo-tagged photos within the i-th group, a is a number of groups, G i is the i-th group, F s is the visual content similarity between the geo-untagged photo and the k-th geo-tagged photo from the set of large-scale samples, n(F s ) is a number of F s (s) , GeoSig i is a Geo-signature of the i-th group and Geo(G i+1 ) is computing the Geo-signature of the next group (the (i+1) -th group).

Geo-signature of each group is used to find the location of geo-untagged photo. Since the geo-signature of each group is defined from the geo-untagged photo. All groups are compared together; resulting in a group has the highest unique evidences about the geo-untagged photo as shown in Fig. 7.

The location of geo-untagged photo is the i-th group that has the highest Geo-signature. From Fig. 8. , Venice has the highest Geo-signature of the geo-untagged photo (GeoSigVenice=250). Hence, the geo-untagged photo was taken from Venice, Italy which can be estimated by Eq. (8).

where i is a sequence of the i-th group, GeoSig i is a geo-signature of the i-th group, GeoSig i+1 is a Geo-signature of the next group (the (i+1) –th group) and GeoSig i+2 is a geo-signature of the next of next group (the (i+2) –th group).

Fig 7 Geo-signature of each group that generated from the geo-untagged photo

4. Experiments and Comparisons

4.1 Experimental Design

In this section, the G-sigMR comparisons were made with other researches about location estimation of a photo: IM2GPS [ 3 ], SC [ 4 ], CS [ 5 ], VSA [ 6 ], MSER [ 7 ], and VCG [ 8 ], respectively. The photos used in our experiment were downloaded from GOLD (Geo-tagged large-scale dataset) [ 9 , 10 ] which contains more than 3.3 million geo-tagged photos. It covers more than 65,000 places around the world that partly focus on interesting places in China, Europe and America as shown in Fig. 8. GOLD was designed to be a test set for evaluation of location estimation and place recognition from a single photo (which can be directly downloaded from ).

Fig 8 Some geo-tagged photos from GOLD [9-10].

The experimental setup was carried out under the environment of Intel Core(TM)2, Quad CP Q 8400 and 4 8GB of RAM. Our G-sigMR was implemented by M-script language in Matlab R2015a as shown in Fig 9 The source code can be requested for the performance comparison with our G-sigMR via the email (under in terms of use) For the experiment, the 80 places from GOLD were statistically used Each location, 5,000 photos were randomly selected (as the same condition for evaluation in [8])

Fig 9 Our G-sigMR implementation using M-script of Matlab R2015a

4.2 Criteria Evaluation

Since the randomly selected 5,000 photos for one location can be evaluated by a criteria named “recognition rate” of the i-th location (denoted as RR i). The RR i are computed using TP i, TN i, FP i and FN i. For each location, the RR i was computed by their random selection of 5,000 photos using Eq. (9).

where i is a sequence of the i-th location ( 1 ≤ i ≤ 80 because of 80 places), TP i is a number of photos taken from the i-th location and correctly estimated, TN i is a number of photos not taken from the i-th location and correctly estimated, FP i is a number of photos not taken from the i-th location but wrongly estimated, and FN i is a number of photos taken from the i-th location but wrongly estimated.

In this paper, we used a criteria named “average recognition rate” ( AVG(RR i )) of 80 places as a criteria for G-sigMR comparison with other researches. The AVG(RR i ) can be computed from Eq. (10). , where L is number of locations which equals as 80.

The performance comparison between our G-sigMR and other methods for estimating the location based on 3.3 million geo-tagged photos of GOLD [ 9 , 10 ] is shown in Fig. 10.

Fig 10 Flow of performance evaluation and comparison

4.3 Results and Discussions

The average recognition rate ( AVG(RR i ) ) of different methods are illustrated in Fig. 11. From the results, our G-sigMR produced more correctness in term of AVG(RR i ) which equals as 94.17.

Fig 11 The AVG(RRi ) of IM2GPS [ 3], SC [ 4], CS [ 5], MSER [ 6], VSA [ 7], VCG [ 8] and our G-sigMR using GOLD [ 9, 10]

Traditional approaches that directly estimate the location from a photo such as IM2GPS [ 3 ] and SC [ 4 ] were lower than 60%. The visual synonym approaches are VSA [6] ( AVG(RR i ) = 85.01), MSER [ 7 ] ( AVG(RR i ) = 85.47) and VCG [ 8 ] ( AVG(RR i ) = 88.16) which obviously can help for improvement of the recognition performance. However, they need a summarization of unique evidence from one place that to be distinguished from other places. Moreover, the structure of indexing is also useful for location estimation of a photo (as CS [ 5 ]). The G-sigMR is able to correctly estimate more than 90%. Since the workflow defines the uniqueness of a location named Geo-signature based on the visual synonym. Its Map-Reduce structure is used to index the similar scenes from million geo-taggedphotos from the set of large-scale samples.

5. Conclusion

In this paper, we have designed the G-sigMR (Geo-signature MapReduce) workflow for location estimation of a photo. The previous approaches are not enough for defining the uniqueness of a place. However, G-sigMR could generate the uniqueness named Geo-signature of a place with the MapReduce structure suitable for large-scale dataset. In the experiment, we used 3.3 million geo-tagged photos from the Geo-tagged large-scale dataset (GOLD) which covers more than 65,000 places in China, Europe and America. G-sigMR is implemented using M-script in Matlab 2015a. For the evaluation, G-sigMR is also compared to other approaches: IM2GPS, SC, CS, MSER, VSA and VCG. In each test, the 80 places are statistically used. Each location is randomly selected 5,000 photos. From the results, G-sigMR produces a better performance in term of average recognition rate that is satisfactory (above 90%). The G-sigMR can be useful in modern applications, i.e., in social multimedia as a new function for location estimation from social photos; a search-based tourism application for finding the family photos of a place; and some forensic applications such as criminal location investigation from the scene for polices [ 54 ].

For future work, the direction of our G-sigMR will combine with the textual metadata such as annotations, tags, duplicated comments, and users’ post from the social multimedia. Textual metadata will be filtered and summarized as the unique words or phrases of the geographical location by deep learning. Since the photo location estimation with the help of textual metadata has the challenge about information correctness that arbitrarily shared from social users. Textual metadata will be work for photo location estimation if information available on the social multimedia covers sufficiently all places around the world.

References

  1. The Digital Universe Rich Data and the Increasing Value of the Internet of Things Zwolenski Matt, Weatherill Lee. Australian Journal of Telecommunications and the Digital Economy.2014-oct.
  2. When Location Meets Social Multimedia Ji Rongrong, Gao Yue, Liu Wei, Xie Xing, Tian Qi, Li Xuelong. ACM Transactions on Intelligent Systems and Technology.2015-mar;:1-18.
  3. IM2GPS: estimating geographic information from a single image Hays James, Efros AlexeiA. 2008 IEEE Conference on Computer Vision and Pattern Recognition.2008.
  4. Spatial coding for large scale partial-duplicate web image search Zhou Wengang, Lu Yijuan, Li Houqiang, Song Yibing, Tian Qi. Proceedings of the international conference on Multimedia - MM \textquotesingle10.2010.
  5. GPS Estimation for Places of Interest From Social Users\textquotesingle Uploaded Photos Li Jing, Qian Xueming, Tang YuanYan, Yang Linjun, Mei Tao. IEEE Transactions on Multimedia.2013-dec;:2058-2071.
  6. Bundling features for large scale partial-duplicate web image search Wu Zhong, Ke Qifa, Isard M, Sun Jian. 2009 IEEE Conference on Computer Vision and Pattern Recognition.2009.
  7. Visual word spatial arrangement for image retrieval and classification Penatti OtávioAB, Silva FernandaB, Valle Eduardo, Gouet-Brunet Valerie, Torres RicardodaS. Pattern Recognition.2014-feb;:705-720.
  8. Image Location Estimation by Salient Region Matching Qian Xueming, Zhao Yisi, Han Junwei. IEEE Transactions on Image Processing.2015-nov;:4348-4358.
  9. GPS Estimation from Users’ Photos Li Jing, Qian Xueming, Tang YuanYan, Yang Linjun, Liu Chaoteng. Lecture Notes in Computer Science.2013;:118-129.
  10. Enhancing Sketch-Based Image Retrieval by Re-Ranking and Relevance Feedback Qian Xueming, Tan Xianglong, Zhang Yuting, Hong Richang, Wang Meng. IEEE Transactions on Image Processing.2016-jan;:195-208.
  11. Unsupervised Image Classification by Probabilistic Latent Semantic Analysis for the Annotation of Images Olaode AbassA, Naghdy Golshah, Todd CatherineA. 2014 International Conference on Digital Image Computing: Techniques and Applications (DICTA).2014.
  12. Using Multi-Descriptors for Real Time Cosmetic Image Retrieval Areeyapinan Jennisa, Kanongchaiyos Pizzanu, Kawewong Aram. Engineering Journal.2014-oct;:97-111.
  13. Detecting Image Forgery Using XOR and Determinant of Pixels for Image Forensics Mookdarsanit Pakpoom, Soimart Lawankorn, Ketcham Mahasak, Hnoohom Narit. 2015 11th International Conference on Signal-Image Technology & Internet-Based Systems (SITIS).2015.
  14. Efficient region of interest detection using blind image division Olaode AbassA, Naghdy Golshah, Todd CatherineA. 2015 Signal Processing Symposium (SPSympo).2015.
  15. Bag-of-Visual Words Codebook Development for the Semantic Content Based Annotation of Images Olaode AbassA, Naghdy Golshah, Todd CatherineA. 2015 11th International Conference on Signal-Image Technology & Internet-Based Systems (SITIS).2015.
  16. Color image segmentation using mean shift filtering and wavelet frames Ozden M, Polat E, Tuna E. Proceedings of the IEEE 13th Signal Processing and Communications Applications Conference, 2005...
  17. An Efficient Algorithm for Earth Surface Interpretation from Satellite Imagery Soimart Lawankorn, Ketcham Mahasak. Engineering Journal.2016-nov;:215-228.
  18. TagPix: Automatic Real-Time Landscape Photo Tagging for Smartphones Debnath Hillol, Borcea Cristian. 2013 International Conference on MOBILe Wireless MiddleWARE, Operating Systems, and Applications.2013.
  19. Landmark classification in large-scale image collections Li Yunpeng, Crandall DavidJ, Huttenlocher DanielP. 2009 IEEE 12th International Conference on Computer Vision.2009.
  20. Image Search with Selective Match Kernels: Aggregation Across Single and Multiple Images Tolias Giorgos, Avrithis Yannis, Jégou Hervé. International Journal of Computer Vision.2015-mar;:247-261.
  21. City-scale landmark identification on mobile devices Chen DavidM, Baatz Georges, Koser Kevin, Tsai SamS, Vedantham Ramakrishna, Pylvanainen Timo, Roimela Kimmo, Chen Xin, Bach Jeff, Pollefeys Marc, Girod Bernd, Grzeszczuk Radek. CVPR 2011.2011.
  22. Visual Place Recognition with Repetitive Structures Torii Akihiko, Sivic Josef, Pajdla Toma, Okutomi Masatoshi. 2013 IEEE Conference on Computer Vision and Pattern Recognition.2013.
  23. Query Specific Rank Fusion for Image Retrieval Zhang Shaoting, Yang Ming, Cour Timothee, Yu Kai, Metaxas DimitrisN. IEEE Transactions on Pattern Analysis and Machine Intelligence.2015-apr;:803-815.
  24. Object recognition from local scale-invariant features Lowe DG. Proceedings of the Seventh IEEE International Conference on Computer Vision.1999.
  25. Distinctive Image Features from Scale-Invariant Keypoints Lowe DavidG. International Journal of Computer Vision.2004-nov;:91-110.
  26. Shape indexing using approximate nearest-neighbour search in high-dimensional spaces Beis JS, Lowe DG. Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition..
  27. Speeded-Up Robust Features (SURF) Bay Herbert, Ess Andreas, Tuytelaars Tinne, Gool LucVan. Computer Vision and Image Understanding.2008-jun;:346-359.
  28. Histograms of Oriented Gradients for Human Detection Dalal N, Triggs B. 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR\textquotesingle05)..
  29. A performance evaluation of gradient field HOG descriptor for sketch based image retrieval Hu Rui, Collomosse John. Computer Vision and Image Understanding.2013-jul;:790-806.
  30. Descriptive visual words and visual phrases for image applications Zhang Shiliang, Tian Qi, Hua Gang, Huang Qingming, Li Shipeng. Proceedings of the seventeen ACM international conference on Multimedia - MM \textquotesingle09.2009.
  31. Visual synonyms for landmark image retrieval Gavves Efstratios, Snoek CeesGM, Smeulders ArnoldWM. Computer Vision and Image Understanding.2012-feb;:238-249.
  32. Effective near-duplicate image retrieval with image-specific visual phrase selection Chen Jiansong, Feng Bailan, Zhu Lei, Ding Peng, Xu Bo. 2012 19th IEEE International Conference on Image Processing.2012.
  33. Lost in quantization: Improving particular object retrieval in large scale image databases Philbin James, Chum Ondrej, Isard Michael, Sivic Josef, Zisserman Andrew. 2008 IEEE Conference on Computer Vision and Pattern Recognition.2008.
  34. Contextual synonym dictionary for visual object retrieval Tang Wenbin, Cai Rui, Li Zhiwei, Zhang Lei. Proceedings of the 19th ACM international conference on Multimedia - MM \textquotesingle11.2011.
  35. Detecting Image Forgery Using XOR and Determinant of Pixels for Image Forensics Mookdarsanit Pakpoom, Soimart Lawankorn, Ketcham Mahasak, Hnoohom Narit. 2015 11th International Conference on Signal-Image Technology & Internet-Based Systems (SITIS).2015.
  36. Remote sensing data processing acceleration based on multi-core processors Zheng Xiao, Xue Yong, Guang Jie, Liu Jia. 2016 IEEE International Geoscience and Remote Sensing Symposium (IGARSS).2016.
  37. Scalable Mobile Image Retrieval by Exploring Contextual Saliency Yang Xiyu, Qian Xueming, Xue Yao. IEEE Transactions on Image Processing.2015-jun;:1709-1721.
  38. Learning salient visual word for scalable mobile image retrieval Yang Xiyu, Qian Xueming, Mei Tao. Pattern Recognition.2015-oct;:3093-3101.
  39. Geolocation based image annotation Shimada Atsushi, Nagahara Hajime, Taniguchi Rin-Ichiro, Charvillat Vincent. The First Asian Conference on Pattern Recognition.2011.
  40. What Is Happening in the Video? —Annotate Video by Sentence Qian Xueming, Liu Xiaoxiao, Ma Xiang, Lu Dan, Xu Chenyang. IEEE Transactions on Circuits and Systems for Video Technology.2016-sep;:1746-1757.
  41. Semantic Annotation of High-Resolution Satellite Images via Weakly Supervised Learning Yao Xiwen, Han Junwei, Cheng Gong, Qian Xueming, Guo Lei. IEEE Transactions on Geoscience and Remote Sensing.2016-jun;:3660-3671.
  42. Multimodal geo-tagging in social media websites using hierarchical spatial segmentation Kelm Pascal, Schmiedeke Sebastian, Sikora Thomas. Proceedings of the 5th International Workshop on Location-Based Social Networks - LBSN \textquotesingle12.2012.
  43. Tag-Based Image Search by Social Re-ranking Lu Dan, Liu Xiaoxiao, Qian Xueming. IEEE Transactions on Multimedia.2016-aug;:1628-1639.
  44. Learning Social Tag Relevance by Neighbor Voting Li Xirong, Snoek CGM, Worring M. IEEE Transactions on Multimedia.2009-nov;:1310-1322.
  45. Rating Prediction Based on Social Sentiment From Textual Reviews Lei Xiaojiang, Qian Xueming, Zhao Guoshuai. IEEE Transactions on Multimedia.2016-sep;:1910-1921.
  46. Service Rating Prediction by Exploring Social Mobile Users’ Geographical Locations Zhao Guoshuai, Qian Xueming, Kang Chen. IEEE Transactions on Big Data.2017-mar;:67-78.
  47. User-Service Rating Prediction by Exploring Social Users\textquotesingle Rating Behaviors Zhao Guoshuai, Qian Xueming, Xie Xing. IEEE Transactions on Multimedia.2016-mar;:496-506.
  48. Placing images on the world map Hauff Claudia, Houben Geert-Jan. Proceedings of the 35th international ACM SIGIR conference on Research and development in information retrieval - SIGIR \textquotesingle12.2012.
  49. Detecting Image Forgery Using XOR and Determinant of Pixels for Image Forensics Mookdarsanit Pakpoom, Soimart Lawankorn, Ketcham Mahasak, Hnoohom Narit. 2015 11th International Conference on Signal-Image Technology & Internet-Based Systems (SITIS).2015.
  50. Personalized Travel Sequence Recommendation on Multi-Source Big Social Media Jiang Shuhui, Qian Xueming, Mei Tao, Fu Yun. IEEE Transactions on Big Data.2016-mar;:43-56.
  51. Efficient data search using map reduce framework Dhekale PratibhaP, Jichkar Nilima. 2016 World Conference on Futuristic Trends in Research and Innovation for Social Welfare (Startup Conclave).2016.
  52. Light-weight operation of a failover system for Cloud computing Mookdarsanit Pakpoom, Gertphol Sethavidh. 2013 5th International Conference on Knowledge and Smart Technology (KST).2013.
  53. A fast map-reduce algorithm for burst errors in big data cloud storage Qin Xue, Kelley Brian, Saedy Mahdy. 2015 10th System of Systems Engineering Conference (SoSE).2015.
  54. To tweet or not to tweet about schizophrenia Bodart Angelique. .2013-dec.
View article in other formats
Author Biographies
Pakpoom Mookdarsanit

Faculty of Information Technology, King Mongkut’s University of Technology North Bangkok, Thailand

Montean Rattanasiriwongwut

Faculty of Information Technology, King Mongkut’s University of Technology North Bangkok, Thailand

Published
Vol 21 No 3, Jun 15, 2017
How to Cite
P. Mookdarsanit and M. Rattanasiriwongwut, “Location Estimation of a Photo: A Geo-signature MapReduce Workflow”, Eng. J., vol. 21, no. 3, pp. 295-308, Jun. 2017.

Authors who publish with Engineering Journal agree to transfer all copyright rights in and to the above work to the Engineering Journal (EJ)'s Editorial Board so that EJ's Editorial Board shall have the right to publish the work for nonprofit use in any media or form. In return, authors retain: (1) all proprietary rights other than copyright; (2) re-use of all or part of the above paper in their other work; (3) right to reproduce or authorize others to reproduce the above paper for authors' personal use or for company use if the source and EJ's copyright notice is indicated, and if the reproduction is not made for the purpose of sale.

Article Statistics
Total PDF downloads: 249
Total XML downloads: 7