e-VDS

We are building and releasing sophisticated video data sets to train machines to recognise objects in our environment. Check the download section below!

Currently, e-VDS35 has 35 classes and a total of 2050 videos of roughly 10 seconds each (see histogram below). We are aiming to collect overall 1750 (50 × 35) videos with your help.

Download

In the section below "raw data" stands for the video as they have been provided to us; no filtering and variable formats. "Processed data" stands for "raw data" which minor sides have been resized to 256 px, videos with less than 144 frames have been removed, clips longer than 564 frames have been trimmed, and everything has been re-encoded in MP4.

• May 2017 version
  
  • Raw data (985 files, 22GB): e-VDS35-May17.tar.
  • Processed data (970 files, 660MB): e-VDS35-May17-m256.tar.
• Raw data from contributors (1035 files, 14GB): FTP directory.

We also provide access to a 33-class image data set (33-image-set.tar) of 300k images, subset of the Open Images data set, fetched with our 9M-URLs crawler. These classes are very similar to the ones found in e-VDS, and each one has roughly 10k training samples and up to 500 validation samples. Therefore, this data set can be used to improve generalisation of your model.

Citation

If you find our data set useful, please cite it as:

@misc{e-VDS,
  author = {Culurciello, Eugenio and Canziani, Alfredo},
  title = {{e-Lab} Video Data Set},
  howpublished = {\url{https://engineering.purdue.edu/elab/eVDS/}},
  year={2017}
}
                    

We greatly appreciate our contributors for having a huge impact on our video data set.

We ask you to record 10-second videos of various household objects you have at home. Please Choose one or more items from the left of the following diagram. These videos are currently the most sought after. Save files in directories with exactly the same names as the categories (you can download this ZIP or TGZ file containing all the empty folders you may need).

• Use your cellphone or camera in landscape mode (horizontally).
• Use a resolution greater or equal to 720p (1280 × 720 pixels).
• Record roughly 10 seconds around the object. Few seconds more is fine.
• Rotate your camera around the object, to get multiple perspectives.
• Keep the object in the center of the image at all times.
• You can use the same object in different environments, but please no more than 5 video for each object.
• Best if you can do ~5 different objects per category.
• Try to record all categories you can at your home / environment.
• No need to place your objects in specific configurations, have high-contrast backgrounds. Record them as you use them in the house.
• Light levels are not important, as long as object can be seen (not completely black or burnt). Use your usual lighting.

Shoot us an e-mail by clicking here or use the following email and subject:

with a link to an online storage service (DropBox, Google Drive, Box, MediFire, Backblaze and similar) so that we can get your clips and add you to the leaderboard.