1

Seven international groups from academia participated in the task and submitted a total of 38 runs. The group with the highest number of submissions had 13 submissions. In the following sections, the methods of the groups are explained (in alphabetical order) and references to further work are given. 3.1

Email: {mbrendel, daroczyb, benczur}@ilab.sztaki.hu The task of object retrieval is to classify objects found in images. This means to find an objects in an image, which is similar to sample objects in the pre-classified images. There are two problems with this task: the first is, how do we model objects. The second is, how do we measure similarity of objects. Our first answer to the first question is to model objects with image segments. Segment, region or blob based image similarity is a common method in content based image retrieval, see for example [ 4, 27, 2, 21 ]. Respectively, the basis of our first method is to find segments on the query image, which are similar to the objects in the pre-classified images. The image is then classified to be in that class, to which we find the most similar segment in the query image.

Image segmentation in itself is a widely researched and open problem itself. We used an image segmenter developed by our group to extract segments from the query images. Our method is based on a graph-based algorithm developed by Felzenszwalb and Huttenlocher [ 11 ]. We implemented a pre-segmentation method to reduce the computational time and use a different smoothing technique. All images were sized to a fixed resolution. Gaussian-based smoothing helped us cut down high frequency noise. Because of the efficiency of OpenCV6 implementation we did not implement either a resizing or a Gaussian-based smoothing algorithm. As pre-segmentation we built a three-level Gaussian-Laplacian pyramid to define initial pixel groups. The original pyramidbased method, which considers the connection between pixels on different levels too, was modified to eliminate the so-called blocking problem. We used brightness difference to measure distance between pixels: dif f Y (P1, P2) = 0.3∗ | RP2 − RP1 | +0.59∗ | GP2 − GP1 | +0.11∗ | BP2 − BP1 | (1)

After pre-segmentation, we had segments of 16 pixels maximum. To detect complex segments, we modified the original graph-based method by Felzenszwalb and Huttenlocher [ 11 ] with an adaptive threshold system using Euclidean distance to prefer larger regions instead of small regions of the image. Felzenszwalb and Huttenlocher defined an undirected graph G = (V, E) where ∀vi ∈ V corresponds to a pixel in the image, and the edges in E connect certain pairs of neighboring pixels. This graph-based representation of the image reduces the original proposition into a graph cutting challenge. They made a very efficient and linear algorithm that yields a result near to the optimal normalized cut which is one of the NP-full graph problems [ 11, 29 ].

6http://www.intel.com/technology/computing/opencv/ Algorithm 1 Algorithm Segmentation (Isrc, τ1, τ2) τ1 and τ2 are threshold functions. Let I2 be the source image, I1 and I0 are the down-scaled images. Let P (x, y, i) be the pixel P (x, y) in the image on level i (Ii). Let G = (V, E) be an undirected weighted graph where ∀vi ∈ V corresponds to a pixel P (x, y). Each edge (vi, vj ) has a non-negative weight w(vi, vj ).

1. Compute M axweight(R) = maxe∈MST (R,E) w(e) for every coherent group of points R where

M ST (R, E) is the minimal spanning tree 2. Compute Co(R) = τ2(R) + M axweight(R) as the measure of coherence between points in R 3. J oin(R1, R2) if e ∈ E exists so w(e) < min(Co(R1), Co(R2)) is true, where R1 T R2 = ∅ and w(e) is the weight of the border edge e between R1 and R2 4. Repeat steps 1,2,3 for every neighboring group (R1, R2) until possible to join two groups Algorithm 1 sometimes does not find relevant parts with low initial thresholds. To find the relevant borders which would disappear with the graph-based method using high thresholds we calculated the Sobel-gradient image to separate important edges from other remainders.

Similarity of complex objects is usually measured on a feature base. This means the the similarity of the objects is defined by the similarity in a certain feature-space.

dist(Si, Oj ) = d(F (Si), F (Oj )) : Si ∈ S, Oj ∈ O where S is the set of segments and O is the set of objects, dist is the distance function of the objects and segments, d is a distance function in the feature space (usually some of the conventional metrics in the n-dimensional real space), F is the function which assigns features to objects and segments. We extracted from the segments features, like mean color, size, shape information, and histogram information. As shape information a 4 × 4 sized low-resolution variant of the segment (framed in a rectangle with background) was used. Our histograms had 5 bins in each channel. Altogether a 35 dimensional, real valued feature-vector was extracted for each of the segments. The same features were extracted for the objects in the pre-classified images taking them as segments. The background and those classes, which were not requested were ignored. The features of the objects were written to a file, with the class-identifiers, which were extracted from the color-coding. This way we obtained a data-base of class samples, containing features of objects belonging to the classes. After this, the comparison of the objects of the pre-classified sample-images and the segments of the query image was possible. We used Euclidean distance to measure similarity. The distance of the query-image Q was computed as: dist(Q) = min dist(Si, Oj ) : Si ∈ S, Oj ∈ O

i,j where S is the set of segments of image Q, O is the set of the pre-classified sample objects. Q is classified to be in the class of that object, which minimizes the distance. The score of an image was computed as:

score(Q) = 1000/dist(Q) where Q is the query image. (2) (3) (4) In our first method (see budapest-acad315) we found that our segments are much smaller than the objects in the pre-segmented images. It would have been possible to get larger segments by adjusting the segmentation algorithm, however this way we would not get objects, which were really similar to the objects. We found that our segmentation algorithm could not generate segments similar to the the objects in the pre-classified images with any settings of the parameters. Even, if we tried our algorithm on the sample-images, and the segments were approximately of the same size, the segments did not match the pre-classified objects. The reason for this is that pre-segmentation was made by humans and algorithmic segmentation is far from capable of the same result. For example, it is almost impossible to write an algorithm, which would segment a shape of a human being as one segment, if his clothes are different. However, people were one of the classes defined, and the sample images contained people with the entire body as one object. Therefore we modified our method. Our second method is still segment-based. But we also do a segmentation on the sample-images. We took the segmented sample-images, and if a segment was 80% inside of an area of a pre-defined object, then we took this segment as a proper sample for that object. This way a set of sample segments were created. After this the method is similar to the previous, the difference is only that we have sample-segments instead of sample objects, but we treat them the same way anyway. The features of the segments were extracted and they were written to a file, with the identifier of the class, which was extracted from the color-codes. After this, the comparison of the segments of the pre-classified images and the query image was possible. We used Euclidean distance again to measure similarity. The closest segment of the image to a segment in any of the objects was searched.

dist(Q) = min dist(Si, Sj ) : Si ∈ S, Sj ∈ O i,j (5) where S is the segments of image Q, O is the set of segments belonging to the pre-classified objects. The image was classified according to the object, to which the closest segment belongs. As we expected, this modification made the algorithm better. 3.2

HUTCIS: Conventional Supervised Learning using Fusion of Image Features

Email: firstname.lastname@tkk.fi All our 13 runs identified with prefix HUTCIS implement a similar general system architecture with three system stages: 1. Extraction of a large number of global and semi-global image features. Here we interpret global histograms of local descriptors as one type of global image feature. 2. For each individual feature, conventional supervised classification of the test images using the VOC2006 trainval images as the training set.

By using this architecture, we knowingly ignored the aspect of qualitatively different training and test data. The motivation was to provide a baseline performance level that could be achieved by just applying a well-working implementation of the conventional supervised learning approach. Table 2 with ROC AUC performances in VOC 2006 test set reveals that the performance of our principal run HUTCIS SVM FULLIMG ALL is relatively close to the best performances in the last year’s VOC evaluation [ 9 ]. The last row of the table indicates what the rank of the HUTCIS SVM FULLIMG ALL run would have been among the 19 VOC 2006 participants.

The following briefly describes the components of the architecture. For more detailed description, see e.g. [ 34 ].

Features For different runs, the features are chosen from a set of feature vectors, each with several components. Table 3 lists 10 of the features. Additionally, the available feature set includes interest point SIFT feature histograms with different histogram sizes, and concatenations of pairs, triples and quadruples of the tabulated basic feature vectors. The SIFT histogram bins have been selected by clustering part of the images with with self-organising map (SOM) algorithm. Classification and fusion The classification is performed either by a C-SVC implementation built around LIBSVM support vector machine (SVM) library [ 3 ], or a SOM-based classifier [ 19 ]. The SVM classifiers (prefix HUTCIS SVM) are fused together using an additional SVM layer. For the SOM classifiers (prefix HUTCIS PICSOM), the fusion is based on summation of normalized classifier outputs.

The different runs Our principal run HUTCIS SVM FULLIMG ALL realizes all the three system stages in the best way possible. Other runs use subsets of those image features, inferior algorithms or are otherwise predicted to be suboptimal.

The run HUTCIS SVM FULLIMG ALL performs SVM-classification with all the tabulated features, SIFT histograms and and twelve previously hand-picked concatenations of the tabulated features, selected on basis of SOM classifier performance in the VOC2006 task. The runs HUTCIS SVM FULLIMG IP+SC and HUTCIS SVM FULLIMG IP+SC are otherwise similar but use just subsets of the features: SIFT histograms and color histogram, or just SIFT histograms, respectively.

The runs identified by prefix HUTCIS SVM BB are half-baked attempts to account for the different training and test image distributions. These runs are also based on SIFT histogram and color histogram features. For the training images, the features are calculated from the bounding boxes specified in VOC2006 annotations. For the test images, the features are calculated for whole images. The different runs with this prefix correspond to different ways to select the images as a basis for SIFT codebook formation.

The run HUTCIS FULLIMG+BB is the rank based fusion of features extracted from full images and bounding boxes. The runs HUTCIS PICSOM1 and HUTCIS PICSOM2 are otherwise identical but use a different setting of SOM classifier parameter. HUTCIS PICSOM2 smooths less Run id.

FULLIMG ALL FULLIMG IP+SC FULLIMG IP Best in VOC2006 Rank the feature spaces, the detection is based on more local information. Both the runs are based on the full set of features mentioned above. concatenations of pairs, triples and Results The tabulated MAP results of HUTCIS runs are corrupted by our own slip in ordering of the queries that prevented us from fully participating the pool selection for query topics 4–10. For the normal pooling, our mistake excluded us completely, for additional relevance judgments only partly. This is expected to degrade our results quite much for the normal pooling and somewhat less for the additional pool, as participating the pool selection is known to be essential for obtaining good performance figures.

As expected, HUTCIS SVM FULLIMG ALL turned out to be the best of our runs for the uncorrupted query topics 1–3. The idea of extracting features only from bounding boxes of objects in order to reduce the effect of different backgrounds in the training and test images seemed usually not to work as well as features extracted from full images, although results were not very conclusive. Partly this is could be to the asymmetry in feature extraction: the features extracted from bounding boxes of the training objects were compared with the features of whole of the test images. The results of the SOM classifier runs did not provide information that would be of general interest, besides confirming the previously known result of SOM classifiers being inferior to SVMs. 3.3

Email: {hmarinc,hugojair,carloshg}@ccc.inaoep.mx The TIA research group at INAOE, Mexico proposed two methods based on image labeling. Automatic image annotation methods were used for labeling regions within segmented images, and then we performed object retrieval based on the generated annotations. Results were not what we expected, though it can be due to the fact that annotations were defined subjectively and that not enough images were annotated for creating the training set for the annotation method. Two approaches were proposed: a semi-supervised classifier based on unlabeled data and a supervised one, the last method was enhanced with a recent proposed method based on semantic cohesion [ 8 ]. Both approaches followed the following steps: 1. Image segmentation 2. Feature extraction 3. Manual labeling of a small subset of the training set 4. Training a classifier 5. Using the classifier for labeling the test-images 6. Using labels assigned to regions images for object retrieval For both approaches the full collection of images was segmented with the normalized cuts algorithm [ 30 ]. A set of 30 features were extracted from each region; we considered color, shape and texture attributes. We used our own tools for image segmentation, feature extraction and manual labeling [ 22 ]. The considered annotations were the labels of the 10 objects defined for this task. The features for each region together with the manual annotations for each region were used as training set with the two approaches proposed. Each classifier was trained with this dataset and then all of the test images were annotated with such a classifier. Finally, the generated annotations were used for retrieving objects with queries. Queries were created using the labels of the objects defined for this task; and selected as relevant those images with the highest number of regions annotated with the object label. Sample segmented images with their corresponding manual annotations are shown in Figure 4. As we can see the segmentation algorithm works well for some images (isolated cows, close-up of people), however for other objects segmentation is poor (a bicycle, for example). KNN+MRFI, A supervised approach For the supervised approach we used a simple knn classifier for automatically labeling regions. Euclidean distance was used as similarity function. The label of the nearest neighbor (in the training set) for each test-region was assigned as annotation for this region. This was our baseline run (INAOE-TIA-INAOE-RB-KNN ).

The next step consisted of improving annotation performance of knn using an approach called (MRFI ) [ 8 ] which we recently proposed for improving annotation systems. This approach consists of modeling each image (region-annotations pairs) with a Markov random field (MRF ), introducing semantic knowledge, see Figure 5. The top−k more likely annotations for each region are considered. Each of these annotations have a confidence weight related to the relevance of the label to being the correct annotation for that region, according to knn. The MRFI approach uses the relevance weights with semantic information for choosing a unique (the correct) label for each region. Semantic information is considered in the MRF for keeping coherence among annotations assigned to regions within a common image; while the relevance weight is considered for taking into account the confidence of the annotation method (k − nn) on each of the labels, see Figure 5. The (pseudo) optimal configuration of regions-annotations for each image is obtained by minimizing an energy function defined by potentials. For optimization we used standard simulated annealing.

The intuitive idea of the MRFI approach is to guarantee that the labels assigned to regions are coherent among them, taking into account semantic knowledge and the confidence of the annotation system. In previous work, semantic information was obtained from cooccurrences of labels on an external corpus. However for this work semantic association between a pair of labels is given by the normalized number of relevant documents returned by GoogleR to queries generated using the pair of labels. This run is named INAOE-TIA-INAOE-RB-KNN+MRFI, see [ 8 ] for details.

SSAssemble: Semi-supervised Weighted AdaBoost The semi-supervised approach consist of using a recently proposed ensemble of classifiers, called WSA [ 22 ]. Our WSA ensemble uses naive Bayes as its base classifier. A set of these is combined in a cascade based on the AdaBoost technique [ 13 ]. Ensemble methods work by combining a set of base classifiers in some way, such as a voting scheme, producing a combined classifier which usually outperforms a single classifier. When training the ensemble of Bayesian classifiers, WSA considers the unlabeled images on each stage. These are annotated based on the classifier from the previous stage, and then used to train the next classifier. The unlabeled instances are weighted according to a confidence measure based on their predicted probability value; while the labeled instances are weighted according to the classifier error, as in standard AdaBoost. Our method is based on the supervised multi-class AdaBoost ensemble, which has shown to be an efficient scheme to reduce the rate error of different classifiers.

Formally WSA algorithm receives a set of labeled data (L) and a set of unlabeled data (U ). An initial classifier N B1 is build using L. The labels in L are used to evaluate the error of N B1. As in AdaBoost the error is used to weight the examples, increasing the weight of the misclassified examples and keeping the same weight of the correctly classified examples. The classifier is used to predict a class for U with certain probability. In the case of U , the weights are multiplied by the predicted probability of the majority class. Unlabeled examples with high probability of their predicted class will have more influence in the construction of the next classifier than examples with lower probabilities. The next classifier N B2 is build using the weights and predicted class of L ∪ U . N B2 makes new predictions on U and the error of N B2 on all the examples is used to re–weight the examples. This process continues, as in AdaBoost, for a predefined number of cycles or when a classifier has a weighted error greater or equal to 0.5. As in AdaBoost, new instances are classified using a weighted sum of the predicted class of all the constructed base classifiers. WSA is described in algorithm 2.

We faced several problems when performing the annotation image task. The first one was that the training set and the test set were different, so this caused a classification with high error ratio. The second one was due the segmentation algorithm. The automatic segmentation algorithm did not perform well for all images leading to have incorrect segmentation of the objects in the images. The last one concerns to the different criteria for manual labeling of the training set. Due all these facts we did not get good results. We hope improving the annotation task by changing part of the labeling strategy. 3.4

Two approaches were adopted by Microsoft Research Asia (MSRA) to perform the object retrieval task in ImageCLEF 2007. One is based on the visual topic model (VTM); the other is the visual language modelling (VLM) method [ 35 ]. VTM represents an image by a vector of probabilities that the image belongs to a set of visual topics, and categorizes images using SVM classifiers. VLM represents an image as a 2-D document consisting of visual words, trains a statistical language model for each image category, and classifies an image to the category that generates the image with the highest probability. Probabilistic Latent Semantic Analysis (pLSA) [ 12 ], which is a generative model from the text literature, is adopted to find out the visual topics from training images. Different from traditional pLSA, all training images of 10 categories are put together in the training process and about 100 visual topics are discovered finally.

The training process consists of five steps: local feature extraction, visual vocabulary construction, visual topic construction, histogram computation, and classifier training. At first, salient image regions are detected using scale invariant interest point detectors such as the Harris-Laplace and the Laplacian detectors. For each image, about 1,000 to 2,000 salient regions are extracted. Those regions are described by the SIFT descriptor which computes a gradient orientation histogram within the support region. Next, 300 local descriptors are randomly selected from each category and combined together to build a global vocabulary of 3,000 visual words. Based on the vocabulary, images are represented by the frequency of visual words. Then, pLSA is performed to discover the visual topics in the training images. pLSA is also applied to estimate how likely an image belongs to each visual topic. The histogram of the estimated probabilities is taken as the feature representation of that image for classification. For multi-class classification problem, we adopt the one-against-one scheme, and train an SVM classifier with RBF kernel for each possible pair of categories. 3.4.2

VLM: Visual Language Modeling The approach consists of three steps: image representation, visual language model training and object retrieval. Each image is transformed into a matrix of visual words. First, an image is simply segmented into 8 x 8 patches, and the texture histogram feature is extracted from each path. Then all patches in the training set are grouped into 256 clusters based on their features. Next, each path cluster is represented using an 8-bit hash code, which is defined as the visual word. Finally, an image is represented by a matrix of visual words, which is called visual document.

Visual words in a visual document are not independent to each other, but correlated with other words. To simply the model training, we assume that visual words are generated in the order from left to right, and top to bottom and each word is only conditionally dependent on its immediate top and left neighbors, and train a trigram language model for each image category. Given a test image, it is transformed into a matrix of visual words in the same way, and the probability that it is generated by each category is estimated respectively. Finally, the image categories are ranked in the descending order of these probabilities. 3.5

Extraction of multi-channel based interest points An extension of the intensity-based Harris detector [ 16 ] is proposed in [ 25 ]. Because of common photometric variations in imaging conditions such as shading, shadows, specularities and object reflectance, the components of the RGB color system are correlated and therefore sensitive to illumination changes. However, in natural images, high contrast changes may appear. Therefore, a color Harris detector in RGB space does not dramatically change the position of the corners compared to a luminance based approach. Normalized rgb overcomes the correlation of RGB and favors color changes. The main drawback, however, is its instability in dark regions. We can overcome this by using quasi invariant color spaces.

The approach PRIP-PRIP HSI ScIvHarris uses the HSI color space [ 32 ], which is quasiinvariant to shadowing and specular effects. Therefore, changes in lighting conditions in images should not affect the positions of the interest points, resulting in more stable locations. Additionally, the HSI color space discriminates between luminance and color. Therefore, much information can be discarded, and the locations get more sparse and distinct.

The PRIP cbOCS ScIvHarris approach follows a different idea. As proposed in [ 33 ], colors have different occurrence probabilities and therefore different information content. Therefore, rare colors are regarded as more salient than common ones. We wish to find a boosting function so that color vectors having equal information content have equal impact on the saliency function. This transformation can be found analyzing the occurrence probabilities of colors in large image databases. With this change of focus towards rare colors, we aim to discard many repetitive locations and get more stable results on rare features.

The characteristic scale of an interest point is chosen by applying a principal component analysis (PCA) on the image and thus finding a description for the correlation of the multi-channel information [ 31 ]. The characteristic scale is decided when the Laplacian of Gaussian function of this projection and the Harris energy is a maximum at the same location in the image. The final extraction of these interest points and corresponding scales is done by preferring locations with high Harris energy and huge scales. A maximum number of 300 locations per image has been defined, as over-description diminish the overall recognition ability dramatically. Local descriptions of interest points The scale invariant feature transform (SIFT) [ 20 ] showed to give best results in a broad variety of applications [ 23 ]. We used the areas of the extracted interest points as a basis for the description phase. SIFT are basically sampled and normalized gradient histograms, which can lead to multiple descriptions per location. This occurs if there is more than one direction of the gradients regarded as predominant.

Estimating the signature of an image In this bag of visual features approach [ 36 ], we cluster the descriptions of one image to a fixed number of 40 clusters using k-means. The centroids and the proportional sizes of the clusters build the signature of one image having a fixed dimensionality of 40 by 129.

Classification The Earth Mover’s Distance (EMD) [ 28 ] showed to be a suitable metric for comparing image signatures. It takes the proportional sizes of the clusters into account, which seems to gain a lot of discriminative power. The classification itself is done in the most straight forward way possible: for every object category, the smallest distances to another signature build the classification. 3.7

Email: surname@cs.rwth-aachen.de The approach used by the Human Language Technology and Pattern Recognition group of the RWTH Aachen University, Aachen, Germany, to participate in the PASCAL Visual Object Classes Challenge consists of four steps:

2. clustering 3. creation of histograms 4. training of a log-linear model where the first three steps are feature extraction steps and the last is the actual classification step. This approach was first published in [ 6, 7 ].

The method follows the promising approach of considering objects to be constellations of parts which offers the immediate advantages that occlusions can be handled very well, that the geometrical relationship between parts can be modelled (or neglected), and that one can focus on the discriminative parts of an object. That is, one can focus on the image parts that distinguish a certain object from other objects.

The steps of the method are briefly outlined in the following paragraphs. To model the difference in the training and test data, the first three steps have been done for the training and test data individually, and then the according histograms have been extracted for the respective other, so that once the vocabulary was learnt for the training data and once for the test data and the histograms are created for each using both vocabularies. Results, however show that this seems not to be a working approach to tackle divergence in training and testing data. Patch Extraction Given an image, we extract square image patches at up to 500 image points. Additionally, 300 points from a uniform grid of 15×20 cells that is projected onto the image are used. At each of these points a set of square image patches of varying sizes (in this case 7 × 7, 11 × 11, 21 × 21, and 31 × 31 pixels) are extracted and scaled to a common size (in this case 15 × 15 pixels).

In contrast to the interest points from the detector, the grid-points can also fall onto very homogeneous areas of the image. This property is on the one hand important for capturing homogeneity in objects which is not found by the interest point detector and on the other hand it captures parts of the background which usually is a good indicator for an object, as in natural image objects are often found in a “natural” environment.

After the patches are extracted and scaled to a common size, a PCA dimensionality reduction is applied to reduce the large dimensionality of the data, keeping 39 coefficients corresponding to the 40 components of largest variance but discarding the first coefficient corresponding to the largest variance. The first coefficient is discarded to achieve a partial brightness invariance. This approach is suitable because the first PCA coefficient usually accounts for global brightness. Clustering The data are then clustered using a k-means style iterative splitting clustering algorithm to obtain a partition of all extracted patches. To do so, first one Gaussian density is estimated which is then iteratively split to obtain more densities. These densities are then reestimated using k-means until convergence is reached and then the next split is done. It has be shown experimentally that results consistently improve up to 4096 clusters but for more than 4096 clusters the improvement is so small that it is not worth the higher computational demands. Creation of Histograms Once we have the cluster model, we discard all information for each patch except its closest corresponding cluster center identifier. For the test data, this identifier is determined by evaluating the Euclidean distance to all cluster centers for each patch. Thus, the clustering assigns a cluster c(x) ∈ {1, . . . C} to each image patch x and allows us to create histograms of cluster frequencies by counting how many of the extracted patches belong to each of the clusters. The histogram representation h(X) with C bins is then determined by counting and normalization such that hc(X) = L1X PlL=X1 δ(c, c(xl)), where δ denotes the Kronecker delta function, c(xl) is the closest cluster center to xl, and xl is the l-th image patch extracted from image X, from which a total of LX patches are extracted.

Training & Classification Having obtained this representation by histograms of image patches, we define a decision rule for the classification of images. The approach based on maximum likelihood of the class-conditional distributions does not take into account the information of competing classes during training. We can use this information by maximizing the class posterior probability QkK=1 QnN=k 1 p(k|Xkn) instead. Assuming a Gaussian density with pooled covariances for the class-conditional distribution, this maximization is equivalent to maximizing the parameters of a log-linear or maximum entropy model p(k|h) =

1 Z(h) exp

C ! αk + X λkchc , c=1 where Z(h) = PK c=1 λkchc is the renormalization factor. We use a modified k=1 exp αk + PC version of generalized iterative scaling. Bayes’ decision rule is used for classification. 4