Public Resources: Content Sets and Software
Content Sets
- P365LLds: A Places365 Lifelogging version Dataset. (available on-demand; to be available on-line soon)
The task of scene recognition has been classically evaluated using still images representing scenes. In the context of the MobiNetVideo project, we have created a new dataset that extrapolates Places365's classes to lifelogging/egocentric videos. The dataset is made up of 450 videos recorded with smartphones, go-pro and handheld cameras. Videos have been obtained by downloading YouTube videos licensed as Creative Commons. For each scene class in Places365, we include between one (90% of the classes) and four videos. The average length of the videos is 638 frames (around twenty-one seconds) and the median length is 600 frames per video (around twenty seconds). In overall, the dataset is approximately 34.1 GB large.
- USSds: A Unified Semantic Segmentation Dataset (available on-demand; to be available on-line soon)
There is a large variety of semantic datasets. However, not all of them have the same semantic classes, and the appearance of shared classes substantially differ. The USSds represents a data integration effort to create a unified semantic dataset which—by enlarging the number of classes and the diversity of the shared classes, aims to provide a more generic benchmark for training and evaluation. The merged datasets have been relabelled to a common set of 293 semantic labels distributed into a total of 145,555 training images and 7,614 validation images. The datasets agglutinated to compose the USSds dataset are: COCO-Stuff Dataset, Cityscapes Dataset, ADE20K Dataset, TASKONOMY Dataset, and Mapillary Dataset.
Software
- Pytorch Implementation of Semantic-Aware Scene Recognition (Alejandro López-Cifuentes, Marcos Escudero-Viñolo, Jesús Bescós, Álvaro García-Martín, "Semantic-aware scene recognition, Pattern Recognition", Volume 102, June 2020, 107256, ISSN 0031-3203, (DOI 10.1016/j.patcog.2020.107256))
- Support for Pytorch and Caffe CNNs Visualization (Learning how to modify training rates in scene-recognition, Miguel Basarte Mena, (advisor: Marcos Escudero Viñolo), Trabajo Fin de Máster (Master Thesis), Master en Investigación e Innovación en TIC – Programa Internacional de Múltiple Titulación IPCV (Image Processing and Computer Vision), Univ. Autónoma de Madrid, Sep. 2019.)
- Pytorch Implementation of "Online Clustering-based Multi-Camera Vehicle Tracking in Scenarios with Overlapping FOVs" (Elena Luna, Juan Carlos San Miguel, José M. Martínez, Marcos Escudero-Viñolo, "Online Clustering-based Multi-Camera Vehicle Tracking in Scenarios with Overlapping FOVs", Multimedia Tools and Applications, 81:7063-7083, Feb. 2002(published online 24 Jan. 2022), Springer, ISSN 1380-7501 (printed version), ISSN 1573-7721 (electronic version) (DOI 10.1007/s11042-022-11923-2)
- Pytorch Implementation of "Detection-aware multi-object tracking evaluation" (Detection-aware multi-object tracking evaluation, Jorge Muñoz Aguado (advisor: Juan Carlos San Miguel), Trabajo Fin de Máster (Master Thesis), Máster Universitario en Deep Learning for Audio and Video Signal Processing, Univ. Autónoma de Madrid, Sep. 2021.)
Last update 27/02/2022