New trends and ideas in visual concept detection: the MIR flickr retrieval evaluation initiative

By Mark J. Huiskes, Bart Thomee, Michael S. Lew. Multimedia Information Retrieval, 2010

403 citations recorded by the source corpus.

Subfields: Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/3cfdd7b0dacef10a1578c0e6ed55aa3f3129a437

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

Enhanced representation and multi-task learning for image annotation (2013, methodology): “One promising approach to transfer knowledge from some classes to others is to share information between object classes on the level of individual features [10, 11, 12, 13, 14] or entire object class models [15, 16].”
Dynamic Match Kernel With Deep Convolutional Features for Image Retrieval (2018, methodology): “Following the previous literature [1], [53], for the holidays dataset, we merged it with MIR Flickr 1M (1 million) images [25], so that the size of the final dataset is 1, 001, 491.”
Efficient On-the-fly Category Retrieval Using ConvNets and GPUs (2014, neutral): “MIRFLICKR-1M [22], [23] is used to augment the data from the PASCAL VOC 2007 test set in our later experiments, and comprises 1M unannotated images (aside from quite noisy image tags).”
Unprocessing Images for Learned Raw Denoising (2018, methodology): “To create our synthetic training data, we start with the 1 million images of the MIR Flickr extended dataset [23], setting aside 5% of the dataset for validation and 5% for testing.”
A Decade Survey of Content Based Image Retrieval Using Deep Learning (2020, background): “2 M 50,000 Object Category Images MS COCO [58] 2014 80 82,783 40,504 Common Object Images MIRFlicker-1M [59] 2010 - 1 M - Scene Images Google Landmarks [60] 2017 15 K ∼1 M - Landmark Images”
Real-time Analysis and Visualization of the YFCC100m Dataset (2015, background): “One of the first Flickr datasets, is the popular MIR Flickr collection from 2008 / 2010 [7, 8].”
Improving Neural Networks with Dropout (2013, methodology): “Method Mean Average Precision % Precision at 50 LDA [8] 0.492 0.754 SVM [8] 0.475 0.758 DBN [22] 0.599 0.867 Autoencoder (based on [15]) 0.600 0.875 DBM [22] 0.623 - DBN with dropout finetuning 0.628 0.891 DBM with dropout finetuning 0.632 0.895 Table.”
DeepHash: Getting Regularization, Depth and Fine-Tuning Right (2015, methodology): “Recall @ 100 Recall @ 100 (HR) ments, we present results on Holidays and UKB data sets, combined with the 1 million MIR-FLICKR distractor data set [45].”
The CLEF 2011 Photo Annotation and Concept-based Retrieval Tasks (2011, methodology): “Both tasks make use of a subset of the MIR Flickr 1 Million image dataset [3].”
Image Labeling on a Network: Using Social-Network Metadata for Image Classification (2012, background): “Our approach also compares favorably to the baselines reported in [10].”
Discriminative Transfer Learning with Tree-based Priors (2013, methodology): “The authors of the dataset [10] provided a high-level categorization of the classes which we use to create an initial tree.”
Multimodal Classification Fusion in Real-World Scenarios (2017, neutral): “compositionality and classification [2], [3], [4], [5], [6].”
Quantifying the Visual Concreteness of Words and Topics in Multimodal Datasets (2018, methodology): “Most directly, these scores enable focusing on “concrete visual concepts” (Huiskes et al., 2010; Chen et al., 2015), by issuing image-search queries could be issued exclusively for concrete concepts during dataset construction.”
CycleISP: Real Image Restoration via Improved Data Synthesis (2020, methodology): “* denotes that these methods use variance stabilizing transform (VST) [40] dataset [31] and split them into a ratio of 90:5:5 for training, validation and testing.”
Content-based visual search learned from social media (2012, result): “One concern of the chapter might be that the number of bi-concepts in our current evaluation is relatively small, when compared to single concept benchmarks [22, 48, 105].”
The New Data and New Challenges in Multimedia Research (2015, background): “The initial 2008 release [5] included 25,000 images with ground truth annotations for 29 concepts, while the 2010 release [6] contained one million images without these annotations, but including visual features.”
Improved Multimodal Deep Learning with Variation of Information (2014, methodology): “Following the experimental protocol [12, 27], we randomly split the labeled data into 15000 for training and 10000 for testing, and used 5000 from training set for validation.”