Alessandro Prest, Vicky Kalogeiton, Christian Leistner, Javier Civera, Cordelia Schmid, Vittorio Ferrari
The YouTube-Objects dataset is composed of videos collected from YouTube by querying for the names of 10 object classes of the PASCAL VOC Challenge. It contains between 9 and 24 videos for each class. The duration of each video varies between 30 seconds and 3 minutes. The videos are weakly annotated, i.e. we ensure that each video contains at least one object of the corresponding class. If you use this dataset, please cite  and .
Dataset release v2.3
This release provides the annotations in PASCAL VOC 2007 format for the same 7,000 bounding box annotations from the YTO v2.2 .
You can download the dataset from here. If you use this release, please cite .
You can explore all annotated frames with the Dataset viewer.
Dataset release v2.2
In this release, we improved the quality of the images by fixing some decompression problems. As demonstrated in , we also have better shot boundaries and we have annotated more bounding boxes (6,975) than the ones contained in v1.0 (1,407).
The dataset contains a total of 720,000 frames. In order to eliminate possible confusion when decoding the videos and in the frame numbering, we release individual video frames after decompression.
This release includes almost 7,000 bounding-box annotations . For evaluation purposes we divided the annotated frames into training and test sets and we release them;in this manner, you are in possession of a perfect copy of the dataset as we used in our experiments . In the training set, we annotated one instance per frame, while in the test set we annotated all instances of the desired object class.
In addition to the videos and the bounding-box annotations, this release also includes several materials from our paper :
- Original videos with the audio tracks.
- Optical flow, as produced by .
- Superpixels, as produced by .
These videos were downloaded from the internet, and may subject to copyright. We don’t own the copyright of the videos and only provide them for non-commercial research purposes.
|README.txt||Description of contents||1 January 2015||6.0KB|
|Ranges.tar.gz||Videos and Shots of the dataset||1 January 2015||13.0KB|
|aeroplane.tar.gz||1 January 2015||1.8GB|
|bird.tar.gz||1 January 2015||3.0GB|
|boat.tar.gz||1 January 2015||11.0GB|
|car.tar.gz||1 January 2015||2.9GB|
|cat.tar.gz||1 January 2015||5.4GB|
|cow.tar.gz||1 January 2015||3.1GB|
|dog.tar.gz||1 January 2015||11.0GB|
|horse.tar.gz||1 January 2015||8.1GB|
|motorbike.tar.gz||1 January 2015||6.0GB|
|train.tar.gz||1 January 2015||11.0GB|
|GroundTruth.tar.gz||Ground truth annotations||7 April 2015||91.0KB|
|OpticalFlow.tar.gz||Optical flow by ||7 April 2015||4.3GB|
|SlicSuperpixels.tar.gz||Superpixels by ||17 April 2015||16.7GB|
|UsefulFiles.tar.gz||Useful files for the dataset||1 January 2015||27.0MB|
|YouTubeObjectsVideos.tar.gz||Videos (including audios)||1 January 2015||6.3GB|
Dataset release v1.0
This release contains a total of 570’000 frames. As demonstrated in , the quality of the video frames play a crucial role in the performance of an object detector trained on them. We release individual video frames after decompression and after shot partitioning. In this manner, you are in possession of a perfect copy of the dataset as we used in our experiments .
In addition to the videos, this release also includes several materials from our paper 
- Bounding-boxes annotations. For evaluation purposes we annotated the object location in a few hundred video frames for each class (see sec. 6.1 ).
- Point tracks and motion segments. As produced by .
- Tubes. Spatio-temporal bounding-boxes as described in section 3.2 . We include all candidate tubes as well as the tube automatically selected by our method.
|code.tar.gz||MATLAB source code to access the Youtube-Objects dataset.||17 June 2012||1MB|
|aeroplane.tar.gz||17 June 2012||2.0GB|
|bird.tar.gz||17 June 2012||3.0GB|
|boat.tar.gz||17 June 2012||7.6GB|
|car.tar.gz||17 June 2012||1.7GB|
|cat.tar.gz||17 June 2012||5.2GB|
|cow.tar.gz||17 June 2012||6.1GB|
|dog.tar.gz||17 June 2012||19.5GB|
|horse.tar.gz||17 June 2012||14.7GB|
|motorbike.tar.gz||17 June 2012||4.3GB|
|train.tar.gz||17 June 2012||21.1GB|
- Learning Object Class Detectors from Weakly Annotated Video
In Computer Vision and Pattern Recognition (CVPR), 2012.
- Object segmentation by long term analysis of point trajectories
In European Conference on Computer Vision (ECCV), 2010.
- Analysing domain shift factors between videos and images for object detection
In PAMI, 2016.
- Large Displacement Optical Flow: Descriptor Matching in Variational Motion Estimation
In PAMI , 2011.
- SLIC Superpixels Compared to State-of-the-art Superpixel Methods
In PAMI , 2012.
This work was partially funded by the QUAERO project supported by OSEO, French State agency for innovation, the European integrated projects AXES and RoboEarth, DPI2009-07130, SNSF IZK0Z2-136096, CAIDGA IT 26/10, a Google Research Award and the ERC projects VisCul and ALLEGRO.