Description
To research and tackle many unique issues in logo detection from live streaming sport events such as multiple instances of the same object, data redundancy, objects appear on the edges of the image and far away from the camera, we contribute a large scale sports video logo detection database called Sports Match Video Logo Dataset (SMVL). Popular basketball and soccer match videos are selected as the data source for this dataset, due to their rich data and frequent logo instances occurrence. The dataset contains videos from 40 live matches, 20 basketball and 20 soccer games. From each match, video clips of around 5 to 10 seconds are extracted from various camera angles. Frame images are extracted from the clips at a rate of 24 frames per second.
Statistics
The dataset contains 199 video clips with around 120 images per clip, resulting in 24,348 frame images and 223 logo classes in total. The resolution of the frames is $3840\times2160$ pixels. For dataset training-evaluation split, 11,979 training split and 12,369 evaluation split are randomly sampled. Another notable feature of this dataset is that each image contains mulitple logo instances and average area of logo instances relative to the whole image is very small, compared with existing popular logo datasets such as FlickrLogos-32, Openlogo and WebLogo-2M.
Dataset Statitics.
| Classes | Training Split | Evaluation Split | Total Images | Average logos per Image | Average logo size |
|---|---|---|---|---|---|
| 223 | 11,979(49.2%) | 12,369(50.8%) | 24,348 | 15.06 | 0.0036% |
Download
SMVL Dataset Images (5.1 GB): [Google Drive]
SMVL Dataset Annotations (18.1 MB): [Google Drive]
SMVL Dataset ImageSets: [Google Drive]
-->Licence
Please notice that, the SMVL Dataset is made available for academic research purpose only. All the images were collected from public sports matches, and the copyright belongs to the original owners.