This document describes the WEBUPV dataset compiled for the ImageCLEF 2015 Scalable Concept Image Annotation challenge. The data mentioned here indicates what is ready for download. However, upon request or depending on feedback from the participants, additional data may be released. The following is the directory structure of the collection, and bellow there is a brief description of what each compressed file contains. The corresponding MD5 checksums of the files shown (for verifying a correct download) can be found in md5sums.txt. Directory structure ------------------- . | |--- README.txt |--- md5sums.txt |--- webupv15_data_lists.zip | |--- annotations/imageclef2015.dev.bbox.*.gz |--- annotations/imageclef2015.dev.textdesc.*.gz |--- annotations/imageclef2015.subtask2cleantrack.*.input_bbox.*.gz |--- annotations/imageclef2015.subtask2cleantrack.dev.textdesc.*.gz |--- annotations/imageclef2015.dev.bbox.supplement.*.gz |--- annotations/imageclef2015.hierarchy.*.gz |--- annotations/imageclef2015.concept_to_parents.*.gz | |--- feats_textual/ | | | |--- webupv15_data_textual_pages.zip | |--- webupv15_data_textual.scofeat.gz | |--- webupv15_data_textual.keywords.gz | |--- feats_visual/ | |--- webupv15_data_visual_images.zip |--- webupv15_data_visual_gist.dfeat.gz |--- webupv15_data_visual_sift_1000.sfeat.gz |--- webupv15_data_visual_csift_1000.sfeat.gz |--- webupv15_data_visual_rgbsift_1000.sfeat.gz |--- webupv15_data_visual_opponentsift_1000.sfeat.gz |--- webupv15_data_visual_colorhist.sfeat.gz |--- webupv15_data_visual_getlf.sfeat.gz |--- webupv15_data_visual_vgg16-fc8.dfeat.gz Contents of files ----------------- * annotations/imageclef2015.dev.bbox.*.gz Development set ground truth localised annotations for sub task 1. The format for the development set of annotated bounding boxes of the concepts is <image_ID> <seq> <Concept> <confidence> <xmin> <ymin> <xmax> <ymax> The development set contains 1,979 images. The bounding boxes may enclose single instances (a single tree) or grouped instances (e.g. group of trees), depending on the context. The annotations are not exhaustive: the emphasis is on concepts that are interesting enough to be described in the image, although background objects are also optionally annotated by our annotators in many cases. Also note that a person might not be annotated if the annotator could not decide whether the person is a man/woman/boy/girl. * annotations/imageclef2015.dev.textdesc.*.gz Development set ground truth textual description annotations of images for sub task 2 The format is: <image_ID> \t <text_description_seq> \t <textual_description> The development set contains 2,000 images with 5 to 51 textual descriptions per image (mean: 9.492, median: 8). Please note that the sentences contain a mix of both American and British English spelling variants (e.g. color vs colour) -- we have decided to retain this variation in the annotations to reflect the challenge of real-world English spelling variants. Basic spell-correction has been performed on the textual descriptions, but we cannot guarantee that they are completely free from spelling or grammatical error. Changelog for v20150306: We have removed some poor textual descriptions that managed to slip past our quality control check. Please use this latest version (v20150306) for your development needs. * annotations/imageclef2015.subtask2cleantrack.*.input_bbox.*.gz Input bounding boxes for the clean track of SubTask 2. This is just a selected subset of 500 development images from imageclef2015.dev.bbox above (please refer to above for file format). * annotations/imageclef2015.subtask2cleantrack.dev.textdesc.*.gz Annotated textual descriptions for 500 development images, to be used to evaluate the content selection ability of the text generation system in the clean track of SubTask 2. The format is the same as the original imageclef2015.dev.textdesc file, except that we further annotated textual terms with their corresponding input bounding boxes, for example [[[dogs|0,4]]] in a textual description refers to the two instances of dogs with the bounding box id 0 and 4 in imageclef2015.subtask2cleantrack.dev.input_bbox. Note that not all descriptions from the original imageclef2015.dev.textdesc are used in this version, and as such the sequence numbers of the descriptions may not necessarily be contiguous as we retained the sequence numbers from the original file for consistency. * annotations/imageclef2015.dev.bbox.supplement.*.gz This file contains all bounding boxes from imageclef2015.dev.bbox, and additional bounding boxes for additional 'general level' categories ('building', 'person', 'animal' etc.) not in the 251 concepts list. These special categories are prefixed with an asterisk(*). We hope that you will find these useful for your development purposes. * annotations/imageclef2015.hierarchy.*.gz The hierarchy structure of the 'general level' categories. File format: *category \t *parent-category \t definition. For example, *mammal is the child of *animal. '#' represents the root node. * annotations/imageclef2015.concept_to_parents.*.gz List of 'general level' category parent(s) for each 251 concept. A concept may have multiple parents (separated by commas). File format: category \t *parent1,*parent2 * annotations/imageclef2015.test.groundtruth.zip Ground truth for the test. * webupv15_data_lists.zip -> data_iids.txt : IDs of the images (IIDs) in the dataset. -> data_rids.txt : IDs of the webpages (RIDs) in the dataset. -> data_*urls.txt : The original URLs from where the images (iurls) and the webpages (rurls) were downloaded. Each line in the file corresponds to an image, starting with the IID and is followed by one or more URLs. -> data_rimgsrc.txt : The URLs of the images as referenced in each of the webpages. Each line of the file is of the form: IID RID URL1 [URL2 ...]. This information is necessary to locate the images within the webpages and it can also be useful as a textual feature. * feats_textual/webupv15_data_textual_pages.zip Contains all of the webpages which referenced the images in the dataset set after being converted to valid xml. In total there are 515754 files, since each image can appear in more than one page, and there can be several versions of same page which differ by the method of conversion to xml. To avoid having too many files in a single directory (which is an issue for some types of partitions), the files are found in subdirectories named using the first two characters of the RID, thus the paths of the files after extraction are of the form: ./WEBUPV/pages/{RID:0:2}/{RID}.{CONVM}.xml.gz To be able to locate the images withing the webpages, the URLs of the images as referenced are provided in the file data_rimgsrc.txt. * feats_textual/webupv15_data_textual.scofeat.gz The processed text extracted from the webpages near where the images appeared. Each line corresponds to one image, having the same order as the data_iids.txt list. The lines start with the image ID, followed by the number of extracted unique words and the corresponding word-score pairs. The scores were derived taking into account 1) the term frequency (TF), 2) the document object model (DOM) attributes, and 3) the word distance to the image. The scores are all integers and for each image the sum of scores is always <=100000 (i.e. it is normalized). * feats_textual/webupv15_data_textual.keywords.gz The words used to find the images when querying image search engines. Each line corresponds to an image (in the same order as in data_iids.txt). The lines are composed of triplets: [keyword] [rank] [search_engine] where [keyword] is the word used to find the image, [rank] is the position given to the image in the query, and [search_engine] is a single character indicating in which search engine it was found ('g':google, 'b':bing, 'y':yahoo). * feats_visual/webupv15_data_visual_images.zip Contains thumbnails (maximum 640 pixels of either width or height) of the images in jpeg format. To avoid having too many files in a single directory (which is an issue for some types of partitions), the files are found in subdirectories named using the first two characters of the IID, thus the paths of the files after extraction are of the form: ./WEBUPV/images/{IID:0:2}/{IID}.jpg * feats_visual/webupv15_*.{s|d}feat.gz The visual features in a simple ASCII text format either in sparse (*.sfeat.gz files) or dense (*.dfeat.gz files). The first line of the file indicates the number of vectors (N) and the dimensionality (DIMS). Then each line corresponds to one vector. For the dense features each line has exactly DIMS values separated by spaces, i.e., the format is: N DIMS Val(1,1) Val(1,2) ... Val(1,DIMS) Val(2,1) Val(1,2) ... Val(2,DIMS) ... Val(N,1) Val(N,2) ... Val(N,DIMS) For the sparse features, each line starts with the number of non-zero elements and is followed by dimension-value pairs, being the first dimension 0, i.e., the format is: N DIMS nz1 Dim(1,1) Val(1,1) ... Dim(1,nz1) Val(1,nz1) nz2 Dim(2,1) Val(2,1) ... Dim(2,nz2) Val(2,nz2) ... nzN Dim(N,1) Val(N,1) ... Dim(N,nzN) Val(N,nzN) The order of the features is the same as in the list data_iids.txt. The procedure to extract the SIFT based features in this subdirectory was conducted as follows. Using the ImageMagick software, the images were first rescaled to having a maximum of 240 pixels, of both width and height, while preserving the original aspect ratio, employing the command: convert {IMGIN}.jpg -resize '240>x240>' {IMGOUT}.jpg Then the SIFT features where extracted using the ColorDescriptor software from Koen van de Sande (http://koen.me/research/colordescriptors). As configuration we used, 'densesampling' detector with default parameters, and a hard assignment codebook using a spatial pyramid as 'pyramid-1x1-2x2'. The number in the file name indicates the size of the codebook. All of the vectors of the spatial pyramid are given in the same line, thus keeping only the first 1/5th of the dimensions would be like not using the spatial pyramid. The codebook was generated using 1.25 million randomly selected features and the k-means algorithm. The GIST features were extracted using the LabelMe Toolbox. The images where first resized to 256x256 ignoring original aspect ratio, using 5 scales, 6 orientations and 4 blocks. The other features colorhist and getlf, are both color histogram based extracted using our own implementation. * webupv15_data_visual_vgg16-relu7.dfeat.gz Contains the 4096 dimensional activations of the relu7 layer of Oxford VGG’s 16-layer CNN model, extracted using the Berkeley Caffe library. More details can be found at https://github.com/BVLC/caffe/wiki/Model-Zoo. * webupv15_data_visual_vgg16-fc8.dfeat.gz Contains the 1000 dimensional activations of the fc8 layer of Oxford VGG’s 16-layer CNN model, extracted using the Berkeley Caffe library. More details can be found at https://github.com/BVLC/caffe/wiki/Model-Zoo. The 1000 dimensions correspond to the 1000 categories in webupv15_data_visual_vgg16-fc8_categories.lst: the higher the value in the 1000d feature, the more likely the corresponding synset appears in the image. Contact ------- For further questions, please contact: Mauricio Villegas <mauvilsa@upv.es> Fei Yan <f.yan@surrey.ac.uk>