EFFICIENT VIDEO SEARCH USING IMAGE QUERIES A. Araujo1, M. Makar2, V. Chandrasekhar3, D. Chen1, S. Tsai1, H. Chen1, R. Angst1 and B.

Transcription

1 EFFICIENT VIDEO SEARCH USING IMAGE QUERIES A. Araujo1, M. Makar2, V. Chandrasekhar3, D. Chen1, S. Tsai1, H. Chen1, R. Angst1 and B. Girod1 1 Stanford University, USA 2 Qualco Inc., USA ABSTRACT We study the challenges of iage-based retrieval when the database consists of videos. This variation of visual search is iportant for a broad range of applications that require indexing video databases based on their visual contents. We present new solutions to reduce storage requireents, while at the sae tie iproving video search quality. The video database is preprocessed to find different appearances of the sae visual eleents, and build robust descriptors. Copression algoriths are developed to reduce syste s storage requireents. We introduce a dataset of CNN broadcasts and queries that include photos taken with obile phones and iages of objects. Our experients include pairwise atching and retrieval scenarios. We deonstrate one order of agnitude storage reduction and search quality iproveents of up to 12% in ean average precision, copared to a baseline syste that does not ake use of our techniques. Index Ters efficient video search, iage-based retrieval 1. INTRODUCTION AND RELATED WORK Visual search has becoe a widely studied topic. However, a large body of visual content, naely videos, cannot be searched visually using today s coercial systes. In this work, we explore probles that arise when one searches videos using iage queries, which presents any practical use cases. Users ight take a picture of a video to obtain inforation related to it. In online education, a user ight want to find the segent in a lecture video where a particular slide is presented. A copany ight want to find all appearances of its logo or a particular product in television broadcasts. The fact that today s coercial visual search systes do not index videos is to a large extent a consequence of the enorous redundancy of video when treated siply as a collection of individual fraes. We address this issue by introducing a ethod that reduces syste s storage requireents draatically while boosting video search quality. Our contributions are the following: (I) A ethod of finding and cobining different appearances of the sae visual eleent, with atheatical justification for both iage pairwise atching and iage retrieval experients deriving ore robust, teporallycoherent descriptions. Our ethod iproves video search quality by up to 12% in ean average precision while achieving one order of agnitude storage savings. Also, our ethod is ore eoryefficient than even a siple 10 ties coarser video search technique, with ean average precision iproveents of 61.4% in this case. These results show that we achieve a better discriinativenessinvariance trade-off for this application. (II) A copression algorith that reduces significantly the storage requireents for descriptors and that, when cobined with other copression techniques, allows for the previously entioned storage savings. (III) We provide a coparison between several different ways of cobining different visual eleent appearances, with ours ranking aong the best. Also, soewhat surprisingly, this coparison shows that keeping a descriptor for only one of the appearances ight har search quality. (IV) We show that tracking keypoints is crucial to obtain better video search quality and further reduced storage needs. Our M. Makar perfored the work while at Stanford University. 3 Institute for Infoco Research, Singapore Queries& Database& Fig. 1: Exaples of query iages and ground truth database fraes fro the CNN2h dataset, introduced in this work. experients show that the new descriptions do not work as well if we siply detect keypoints independently for each frae. Related work. Previous work has explored video search by iage. Chen et. al. [1] introduced a technique to find a video fro iages captured with obile phones. TRECVID s Instance Search task [2] has evaluated systes that find repeated objects, persons or locations in a large video collection [3, 4]. Our ethod differs fro [1, 3, 4] since we explicitly cobine different appearances of the sae feature to construct a ore robust representation which also akes our database storage very copact. Sivic and Zisseran illustrated the effectiveness of the Bag-ofWords fraework with a ovie [5], and enhanced search for regions of video fraes with different tracking ethods [6]. By using a Teporally Coherent Detector (TCD) [7], and atching fraes to join tracks, our approach is reiniscent of [6] s, in the sense that it uses short- and long-range tracking ethods. However, our usage of TCD akes the keypoint tracks ore stable due to systeatic tracking on the canonical patch level copared to [6] s ethod of tracking only when a keypoint is issing. One of our key findings deonstrates that, when using teporal aggregation, TCD keypoints perfor better than keypoints that are detected every frae (as used in [6, 5]). This is shown by superior retrieval and pairwise atching results using a saller storage footprint. Copared to [8, 6, 5], our dataset is uch larger and uch ore challenging: it contains 139 queries, fro photos taken with significant distortions and iages of objects collected fro the web copared to [6, 5, 8] s 6 queries, taken fro regions of fraes in the sae database video. Sivic and Zisseran [5] argue that averaging descriptors in a track iproves signal-to-noise ratio, but without evaluating this idea. We build on top of this idea to introduce a good justification for averaging and deonstrate quantitative iproveent in search quality and database storage when using averaged descriptors. Descriptor averaging was also explored by Takacs et al. [9] in a landark recognition application. Our work differs fro [9] since we work with video search, use teporal coherence for averaging, justify the use of averaging and copare different aggregation options. On the copression side, recent work considered copression of descriptors [10, 11, 12], inverted index structures [13] and locations [14]. In this work, we focus on storage gains which arise naturally

2 when using the sae robust aggregated descriptor for a large nuber of keypoints in the database video fraes. Moreover, we design copression algoriths to encode indices that ake for a large portion of storage needs. 2. FINDING AND AGGREGATING VISUAL ELEMENTS We use the expression visual eleent to indicate iage patches that represent the sae part of a specific visual structure in different fraes. For exaple, two iage patches corresponding to a specific wheel of a specific car in two different fraes represent the sae visual eleent. But two descriptors that represent parts of different visual structures (e.g., a car and a tree) do not represent the sae visual eleent, even if they are close in ters of descriptor distance. Our objective is two-fold. First, to generate a ore robust representation of a visual eleent, by aggregating their appearances throughout the video under different conditions, such as different viewpoints and illuination. Second, to enable search in very large video datasets, by draatically reducing descriptor storage needs. We collect different appearances of visual eleent in a Teporally Aggregated Patch Set (TAPS) t, which we define as a set of patches p i, i {1,..., n }, where n = t. These patches represent visual eleent under different conditions. Each patch is assigned to exactly one TAPS. We represent TAPS t by a TAPS descriptor, φ, which we define as a function of patches that are assigned to t, i.e., φ = f(t ). A TAPS descriptor is the aggregated representation we use for all patches of the sae visual eleent. In the following, we present how to detect and track visual eleents, then show their TAPS descriptors can be coputed Keypoint Detection and Tracking We experient with two different ways of finding keypoints in video fraes. The first approach detects keypoints for each frae independently (ID - Independent Detection), as in [5, 8, 6, 3, 4]. The second option is to use a Teporally Coherent Detector (TCD) [7]. TCD tracks keypoints on the canonical patch level (i.e., it searches for the best location, scale and orientation for the keypoint localization in the subsequent fraes). When using ID, keypoints locations, scales and orientations present significant variations in a single track, due to the sensitivity of the detector. In contrast, TCD generates keypoints whose patches and descriptors are uch ore siilar within the sae track [15]. In order to establish correspondences between tracks that are not contiguous in tie (due to teporary occlusion, for exaple), we use a feature-based ethod for atching iages, using a ratio test followed by a geoetric consistency check, siilar to [16]. This is siilar to the long-range tracking approach of [6]. We use a paraeter, N p, to denote the nuber of fraes within which we allow atching operations. In our experients, we set it to 10 or 50 (equivalent to 1 or 5 seconds, respectively). This iage atching procedure is also used with ID ode to find feature tracks, for siplicity. A TAPS is created containing all different patches within a track, for each of the tracks that are found with this ethod TAPS description After collecting repeated visual eleents into a TAPS, we represent each patch fro a TAPS using the sae TAPS descriptor. We consider different ways to generate a TAPS descriptor: Keep One (KO). We keep one of the patches fro each TAPS, and use it to copute a descriptor. Patch Average (PA). Patches are rotated and scaled according to their keypoint. Then, they are averaged and a descriptor is coputed fro the ean patch. This is equivalent to averaging the gradients of the different patches before descriptor coputation. Miniu Distance (MD). We keep all descriptors of the different patches in a TAPS. To calculate the distance fro a query descriptor to a TAPS, we select the iniu distance between the forer to each of the descriptors of the latter i.e., this ode calculates a best case distance for coparison of a query descriptor and a TAPS. Note that in this case we do not reduce descriptor storage requireents, since we still need to keep each of the original descriptors. Descriptor Average (DA). We extract descriptors fro each patch in a TAPS and average the. This follows naturally for pairwise atching and retrieval scenarios, as follows. Pairwise Matching. Consider atching a query iage against a video frae, where we find putative feature atches using the ratio test fro [16]. Consider atching the query descriptor q to the TAPS t. An intuitive way to capture contributions fro different patches in a TAPS is to consider the expected square distance between q and t. We can use Lowe s ratio test [16] in its square version, by using square distances and a square threshold. Note that t is a set of saples fro the probability distribution for patches generated fro the sae visual eleent. Let U R D (D being the descriptor diensionality) be a rando vector denoting the descriptor of a patch drawn fro the probability distribution sapled by t: E U[ q U 2 ] = E U[q T q 2q T U + U T U] (1a) = q T q 2q T µ + [σi 2 + E Ui [U i] 2 ] (1b) i = q µ 2 + σi 2 (1c) i where µ is the ean descriptor fro t s patches, σ i the standard deviation of coponent i of descriptors of patches fro t and x is the L 2 nor of x. This coputes the dissiilarity between each query descriptor and each video frae s TAPS. Both ters in (1c) can be coputed by estiating and storing, for each TAPS, the ean µ and one other nuber, i σ2 i, a easure of intra-taps variance. Retrieval. Consider a Bag-of-Words (BoW) odel for retrieval, to search a large nuber of video fraes without atching the query to each frae. Consider the construction of a codebook using a vector quantizer. In this case, we want to iniize j dist(xj, ˆxj), where x j represents a descriptor, ˆx j its reproduction vector and dist(.) being a nonnegative distortion easure. We want to use the average square distance between a TAPS and a reproduction vector. For a dataset with M TAPS s, K reproduction vectors and assignents c (with c {1, 2,..., K}, the assignent for TAPS ): dist(u, ˆx c ) = E U [ U ˆx c 2 ] (2a) = µ ˆx c 2 + σi 2 (2b),i where µ represents the ean descriptor fro the -th TAPS and σ i 2 the variance of the i-th coponent of TAPS. The ter,i σ2 i is independent of both the codebook and the assignents. Thus, a codebook trained using the K-eans algorith with the average of descriptors fro the sae visual eleent is equivalent to using the expected square distances introduced in (1c). The assignent of a TAPS to one of the codewords can also be done without regard for the intra-taps variance ter: it will be a constant factor added to the expected square distance between a TAPS and each centroid. Note that we only include a TAPS for codebook training once, instead of once for each patch in the database. This akes K-eans training uch faster, since there are usually one order of agnitude fewer TAPS than descriptors. It also allows for redundancy reoval that ay har codebook training, and decreases quantization noise by aking the different appearances of each visual eleent be quantized to a single codeword. This is related to the work of [17, 18]. Intra-TAPS variance. In practice, the intra-taps variance is sall copared to the distance fro the query feature to the average descriptor, not affecting atching results significantly (as seen in 4.2). To easure its iportance, we collected 4M patch pairs fro query iages and database fraes, and easured i σ2 i / q µ 2. Using TCD with N p = 50, the average ratio is 5.3%. In all cases, the intra-taps variance ter can be discarded, as shown in 4.2.

3 n<1& n& Queries& 11 Database& TAPS TCD 10 TAPS TCD 50 (x$,$y)$=$(269$,$305)$ TAPS$nuber$($$$$$$$)$=$257$ log2(probability) 12 TAPS$nuber$($$$$$$$)$=$257$ Closest$keypoint$ $Difference$$$$$$$$$$=$$$0$ TAPS ID 10 TAPS ID (a) TAPS nubers (b) 1000 (c) Fig. 2: (a) Illustration of the predictive coding algorith. For each keypoint in frae n, the algorith finds the spatially closest keypoint in frae n 1 and encodes the difference in TAPS nubers. (b) Distribution for the 1,000 ost frequent TAPS nubers, for different Np and detection ethods. (c) Exaple of pairs that are succesfully atched using TAPS but not with frae-based descriptors. 3. REDUCING STORAGE REQUIREMENTS The ain coponents for storing our retrieval syste (described in 4.3) are: retrieval structure and its index, features, and their locations. In a visual search application with a database of videos, the storage cost can quickly becoe prohibitive. We use the ethod of [13] to store an inverted index ore efficiently. To store keypoints locations, we use a siple unifor quantizer with step size of 2, followed by Arithetic Coding [19]. In the rest of this section, we discuss how to store descriptors with as sall of a cost as possible. The ain advantage of our approach to reducing storage needs is the fact that we can use a TAPS descriptor to represent ultiple database features. By representing a database feature by TAPS t, we need to store the TAPS nuber,, that indicates which TAPS it corresponds to. We are interested in reducing storage requireents by storing TAPS descriptors and each frae s TAPS nubers, instead of each frae s descriptors. In our dataset, TAPS nubers ake for 32% of total descriptor storage, if using 32-bit unsigned integers. We propose a lossless copression schee to encode these nubers ore efficiently, significantly reducing storage cost. First, note that the distribution of TAPS nubers is far fro unifor Fig. 2(b). This allows for ore efficient encoding. We use an Arithetic Coder to efficiently encode a frae s sequence of TAPS nubers. In the pairwise atching step of our retrieval pipeline, we use descriptors to rerank the results returned by the retrieval structure. In our case, we use TAPS descriptors, instead of the original descriptors. Thus, we need to efficiently access TAPS nubers in order to know which TAPS descriptors to use. However, if we can afford soe delay, we can further reduce storage by the use of predictive coding. We define an Encoding Sequence as a sequence of fraes in which all TAPS nubers fro a given frae are predicted by TAPS nubers fro the previous frae except for the first frae in the sequence. The Encoding Sequence Size (ESS) is dependent on how uch delay one can afford before having access to TAPS nubers of the required frae. The worst-case delay before starting to decode TAPS nubers fro the required frae is T (ESS 1), where T is the decoding delay per frae. There is a trade-off between storage requireents and syste delay. Note that ESS = 1 corresponds to encoding TAPS nubers for each frae individually. Predictive coding of TAPS nubers is done as follows. For each keypoint in frae n, we find the spatially closest keypoint in frae n 1. We then encode the sequence of differences in TAPS nubers with an Arithetic Coder. This process is illustrated in Fig. 2(a). 4. EXPERIMENTS We introduce the CNN2h dataset1, coposed of 2 hours of CNN video. We provide annotated ground truth query results for 139 queries coposed of photos taken with obile phones and tablets 1 It can be accessed at fro displays showing the video (with substantial geoetric and photoetric distortions), along with pictures of objects collected fro the web. We provide 2,951 true and 21,412 false ground-truth atching pairs of query iages and database fraes, creating a pairwise atching experient. The fraes are sapled at 10fps a total of 72,000 fraes. Fig. 1 presents saple queries and soe of their corresponding database fraes. The new dataset is needed since other datasets [5, 8, 6, 2] are either too sall or use queries that are only regions in a video frae (none of the being pictures taken with caeras or clean iages of products). Also, [2] s queries include an indication of its type (object, person or place). Our experients use SIFT [16] detector and descriptor, with a budget of 400 features per frae, selected by highest Difference of Gaussian peaks. Using TCD [7], only 7.3% of fraes had keypoints detected independently all others used keypoint tracking. For copression of TAPS nubers, we used ESS = {1, 10} Tracking and storage Table 1 presents tracking and descriptor storage results and Fig. 2(b) presents the distribution of the ost frequent TAPS nubers. TAPS s tend to capture ore patches as Np increases. The ratio of nuber of TAPS s to nuber of descriptors is sallest for higher Np s and when using TCD. Significant storage savings are obtained using TAPS, copared to storing all descriptors in the database: using TCD, 93.44% with Np = 50. Note the iportance of copressing TAPS nubers: storage is reduced by up to 28.18% copared to storing 4 bytes per nuber. Decoding delays can becoe significant with ESS = 10, but can be avoided if using ESS = 1. Storage needs for different retrieval systes using a Scalable Vocabulary Tree (SVT) [20] with 1M nodes are shown in Table 2. Location coding saves 75% of storage for locations. Inverted index copression obtains 95.37% gains for TAPS with TCD using Np = 50. Overall, we construct a syste to index 72,000 fraes using 412MB, copared to a baseline of 4,003MB savings of 89.7%. This is even ore eory-efficient than using descriptors extracted at 1fps with ID ode, as shown in Table 2, with savings of 20.8% Pairwise Matching In this experient, we atch a query iage to a video frae using SIFT with ratio test [16] and RANSAC using an affine odel. The score is the nuber of inliers fro the estiated odel. We draw an ROC curve coparing the different approaches. The iportant operating points of this curve are between 10 2 and 10 3 False Positive Rates (FPR) typical operating points for applications. We copare the different TAPS description odes, using TCD, in Fig. 3(a), for Np = 50. First, note that the KO ode perfors substantially worse than all others. This is intuitive since it does not take into account different appearances of the sae keypoint. Second, the intra-taps variance ter does not affect uch pairwise atch-

4 TPR TAPS DA Np=50, TCD TAPS MD Np=50, TCD TAPS DA with intra TAPS var. ter, Np=50, TCD TAPS KO Np=50, TCD TAPS PA Np=50, TCD FPR 0.01FPR TCD ID TAPS DA Np=10, TCD TAPS DA Np=10, TCD, quant. locations TAPS DA Np=50, TCD TAPS DA Np=50, TCD, quant. locations TAPS DA Np=10, ID TAPS DA Np=50, ID Pairwise atching syste storage (Bytes) TAPS DA Np=50 ID Retrieval syste storage (Bytes) (a) (b) (c) Fig. 3: (a) Coparison of TAPS description odes, using N p = 50 and TCD keypoints. The iportant operating points are between 10 3 and 10 2 FPRs. (b) Pairwise atching results: TPR and storage requireents at FPR = (c) Retrieval results for different systes copared by storage cost. For each configuration, we vary the SVT s size. Note that the two left-ost curves use copressed inverted index and locations. N p Detection ode Nuber of descriptors Nuber of TAPS s Total storage - all database descriptors Total storage - TAPS with 32-bit TAPS nubers AP Total storage - TAPS with ESS = 1 Total storage - TAPS with ESS = 10 TCD ID ID 1fps TAPS DA Np=10 TCD TAPS DA Np=10 TCD copr. index, loc. TAPS DA Np=50 TCD TAPS DA Np=50 TCD copr. index, loc. TAPS DA Np=10 ID Worst case delay for ESS = 10 (s) 10 TCD 27,299,281 1,834, TCD 27,299,281 1,638, ID 28,793,066 3,566, ID 28,793,066 2,781, Table 1: Tracking and descriptor storage results for different N p s and keypoint detection ethods. Note the significant storage gain when using TAPS with copression of TAPS nubers, copared to storing all descriptors approxiately one order of agnitude when using TCD. Decoding delays are easured on an Intel Xeon 2.4GHz processor. Coponent ID, baseline ID, copressed inv. index and locations TCD, baseline TCD, copressed inv. index and locations TAPS-DA TCD, ESS = 10, N p = 50 TAPS-DA TCD, ESS = 10, N p = 50, copressed inv. index and location ID, baseline, 1 fps SVT Inv. Index Locations Total Table 2: Storage requireents for different retrieval systes using a Scalable Vocabulary Tree with 1M nodes. Our syste indexes video at 10fps. It achieves savings of 89.7% with respect to the baseline and is ore eory-efficient than using descriptors at 1fps with ID ode (right-ost colun). ing perforance. Third, the MD ode perfors coparably to the DA ode so we do not need to store all different visual eleent appearances. Lastly, PA and DA perfor on par. We choose to use DA as it is well-justified for both pairwise atching and retrieval experients, presents best pairwise atching perforance and can be stored with uch reduced requireents. We plot search quality against storage requireents for the different detection odes and N p s, with TAPS using DA description ode, in Fig. 3(b). For this application, no retrieval structures are used, so we only need to store descriptors (or TAPS descriptors and TAPS nubers) and locations. TAPS s allow for 92.76% reduced storage at iproved search quality for TCD with N p = 50 and quantized locations. Note that TCD-based TAPS s require lower storage than ID-based ones as TCD keypoints are ore coherent. The reason why the use of TAPS iproves search quality is better teporal coverage due to ore robust descriptors. Fig. 2(c) shows soe pairs for which perforance is iproved using TAPS difficult atches, since they are near a shot transition. Matching fails for frae-based descriptors, but the (iplicit) teporal constraint iposed by TAPS helps finding ore feature atches Retrieval We eploy a BoW retrieval ethod using an SVT, as in [20]. Two ethods are copared: 1) We train an SVT based on all descriptors of the dataset (siilar to [6, 8], but with larger and ore scalable codebooks), and 2) We train an SVT based on the TAPS descriptors using DA ode. We evaluate retrieval perforance for a range of SVT sizes, fixing the SVT s branch factor to 10, and varying the nuber of levels fro 2 to 6. When querying an iage, we score database fraes using TF-IDF and output the top-ranked ones in a short list that will go through a final pairwise atching stage. Philbin et al. [21] showed that there can be significant loss in search quality by quantizing descriptors for that reason, our retrieval syste uses descriptors in a separate pairwise atching stage (using the procedure described in 4.2). We rerank fraes fro the SVT s short list by the nuber of inliers after geoetric verification. We liit the nuber of pairwise atching operations to 50, to avoid retrieval delays. Fraes that are close in tie tend to be ranked close together in the SVT s short list. Pairwise atching fraes that are too close in tie would be a waste of resources, since fraes close in tie are very siilar, and there is little value in providing results too close in tie. We do not perfor pairwise atching for a given frae if another frae, within 1 second, has already been pairwise atched. For perforance assessent, we use ean average precision (AP). For each query, we calculate the axiu possible nuber of atches in the database by discounting the 1 second window, and calculate AP based on the nuber of correct results in the 50 top-ranked positions after pairwise atching. Fig. 3(c) presents results for the retrieval experient. Again, we observe iproveents by using TAPS due to better teporal coverage, by up to 12% (0.07 AP). When coparing based on storage costs, the advantage of using TAPS is clear, with a reduction of up to 89.7%. Our syste perfors significantly better than if just using descriptors with ID ode at 1fps: it presents 61.4% retrieval iproveent (0.25 AP) with 20.8% lower storage cost. 5. CONCLUSION This work considers video search using iage queries. We use teporally coherent keypoints to build better descriptors, in a atheatically justified anner, iproving video search quality and reducing storage needs. We show that descriptor averaging is aong the best description odes. A new copression algorith reduces significantly storage cost of indices that ake for a large portion of total storage needs. In a retrieval experient, search quality iproves by up to 12% AP while reducing storage cost by up to 89.7%. Lastly, we show that teporally coherent keypoints are essential to achieve best video search quality at the lowest possible storage cost.

5 6. REFERENCES [1] D. Chen, N.-M. Cheung, S. Tsai, V. Chandrasekhar, G. Takacs, R. Vedantha, R. Grzeszczuk, and B. Girod, Dynaic selection of a feature-rich query frae for obile video retrieval, in Proc. ICIP, [2] P. Over, G. Awad, M. Michel, J. Fiscus, G. Sanders, W. Kraaij, A. F. Seaton, and G. Quenot, TRECVID 2013 An Overview of the Goals, Tasks, Data, Evaluation Mechaniss and Metrics, in Proc. TRECVID, [3] C. Schulze and S. Palacio, Retrieving Objects, People and Places fro a Video Collection: TRECVID 12 Instance Search Task, in Proc. TRECVID, [4] A. Bursuc, T. Zaharia, O. Martinot, and F. Preteux, ARTEMIS- UBIMEDIA at TRECVid 2012 : Instance Search Task, in Proc. TRECVID, [5] J. Sivic and A. Zisseran, Video Google: a text retrieval approach to object atching in videos, in Proc. ICCV, [6] J. Sivic, F. Schaffalitzky, and A. Zisseran, Object Level Grouping for Video Shots, IJCV, vol. 67, no. 2, [7] M. Makar, S. Tsai, V. Chandrasekhar, D. Chen, and B. Girod, Interfrae Coding of Canonical Patches for Mobile Augented Reality, in Proc. ISM, [8] J. Sivic and A. Zisseran, Video Google: Efficient visual search of videos, Toward Category-Level Object Recognition, vol. 4170, [9] G. Takacs, V. Chandrasekhar, N. Gelfand, Y. Xiong, W.-C. Chen, T. Bispigiannis, R. Grzeszczuk, K. Pulli, and B. Girod, Outdoors augented reality on obile phone using loxel-based visual feature organization, in Proc. MIR, [10] R. Ji, J. C. L.-Y. Duan, H. Yao, Y. Rui, S.-F. Chang, and W. Gao, Towards Low Bit Rate Mobile Visual Search with Multiple-Channel Coding Categories and Subject, in Proc. ACM Multiedia, [11] C. Strecha, A. Bronstein, M. Bronstein, and P. Fua, LDAHash: Iproved Matching with Saller, PAMI, vol. 34, no. 1, [12] V. Chandrasekhar, G. Takacs, D. Chen, S. Tsai, Y. Reznik, R. Grzeszczuk, and B. Girod, Copressed Histogra of Gradients: A Low-Bitrate Descriptor, IJCV, vol. 96, no. 3, [13] D. Chen, S. Tsai, V. Chandrasekhar, G. Takacs, R. Vedantha, R. Grzeszczuk, and B. Girod, Inverted Index Copression for Scalable Iage Matching, in Proc. DCC, [14] S. Tsai, D. Chen, G. Takacs, V. Chandrasekhar, J. Singh, and B. Girod, Location Coding for Mobile Iage Retrieval, in Inforation Systes Journal, [15] M. Makar, Interfrae Copression of Visual Feature for Mobile Augented Reality, Ph.D. thesis, Departent of Electrical Engineering, Stanford University, [16] D. Lowe, Distinctive Iage Features fro Scale-Invariant Keypoints, IJCV, vol. 60, no. 2, Nov [17] F. Jurie and W. Triggs, Creating Efficient Codebooks for Visual Recognition, in Proc. ICCV, [18] R. Ji, H. Yao, X. Xie, and Q. Tian, Vocabulary Hierarchy Optiization and Transfer for Scalable Iage Search, IEEE Multiedia, vol. 18, no. 3, [19] A. Moffat, R. Neal, and I. Witten, Arithetic coding revisited, ACM Trans. on Inf. Syst., vol. 16, no. 3, [20] D. Nister and H. Stewenius, Scalable Recognition with a Vocabulary Tree, in Proc. CVPR, [21] J. Philbin, M. Isard, J. Sivic, and A. Zisseran, Descriptor learning for efficient retrieval, in Proc. ECCV, 2010.