Overview

The Out the Window (OTW) dataset is a crowdsourced activity dataset containing 5,668 instances of 17 activities from the NIST Activities in Extended Video (ActEV) challenge. These videos are crowdsourced from workers on the Amazon Mechanical Turk using a novel scenario acting strategy, which collects multiple instances of natural activities per scenario. Turkers are instructed to lean their mobile device against an upper story window overlooking an outdoor space, walk outside to perform a scenario involving people, vehicles and objects, and finally upload the video to us for annotation. Performance evaluation for activity classification on VIRAT Ground 2.0 shows that the OTW dataset provides an 8.3% improvement in mean classification accuracy, and a 12.5% improvement on the most challenging activities involving people with vehicles.

Data Annotation

Annotation files are CSV format with schema:

[Video ID, Activity ID, Actor ID, Activity or Object Type, Frame Number, XMin, YMin, XMax, YMax, Labeled]

Video ID: A globally unique ID assigned to each video for each dataset. Each Homes video is located in homes/video/VIDEO_ID.mp4. Each Lots video is located in lots/video/VIDEO_ID.mp4
Activity ID: A unique ID assigned to each activity within a specific dataset (lots or homes).
Actor ID: A unique ID assigned to each actor. If we are not sure of the actor, this will be None.
Activity Type: A label for the activity or object in the annotation
Frame Number: The frame number of the annotation. Frame numbers correspond to the output of extract_frames.py
XMin, YMin, XMax, YMax: The bounding box of the annotation defining the upper left corner (XMin, YMin) and bottom right corner (XMax, YMax) in image coordinates, where X=column index, Y=row index in image coordinates.
Labeled: A boolean indicating whether or not a frame was Human Labeled (True) or Interpolated (False). We used a combination of tracking and linear interpolation to generate bounding boxes in between the start and end frames of annotation by a human annotator.

Example annotations for a single activity from ./homes/annotations.csv:

00000,0,00038,dismounting bike,252,82,1165,255,1586,True
00000,0,00038,person,252,85,1165,211,1446,True
00000,0,00038,bicycle,253,103,1230,250,1458,False
00000,0,00038,dismounting bike,253,85,1165,250,1458,False
00000,0,00038,person,253,85,1165,211,1446,False
00000,0,00038,bicycle,254,103,1229,254,1455,False
00000,0,00038,dismounting bike,254,85,1165,254,1455,False
00000,0,00038,person,254,85,1165,211,1446,False
00000,0,00038,bicycle,255,102,1226,255,1449,False
00000,0,00038,dismounting bike,255,85,1165,255,1449,False
00000,0,00038,person,255,85,1165,211,1446,False

Frames are 0-indexed. Given an Activity Directory (homes or lots) and a Frame Number, the python snippet for the absolute path of an frame image filename is:

img_file_name = os.path.join(".", "homes", "%08d.jpg" % framenum)

Instruction

OTW to DIVA annotation

d_otw_to_diva.json is a JSON dictionary that maps OTW label string to their equivalent DIVA label string. For example:

OTW : DIVA
"carrying (large)":"transport_heavycarry",
"pushing cart":"pull"

Frame extraction

A python3.x script for extracting frames from either the homes or lots datasets.

Usage:

pip3 install imageio imageio-ffmpeg
python3 extract_frames.py homes 10 5
python3 extract_frames.py lots

This export will take a while, and will extract frames to ./homes/frames. This will parallelize the extraction over 10 workers (optionally can be increased or decreased) with the PNG compression level of 5 (1=worst, 9=best).

Citation

Please use the following citation when referencing the dataset:

@article{Castan2019OutTW,
  title={Out the Window: A Crowd-Sourced Dataset for Activity Classification in Surveillance
Video},
  author={Greg Casta{\~n}{\'o}n and Nathan Shnidman and T. Anderson and J. Byrne},
  journal={ArXiv},
  year={2019},
  volume={abs/1908.10899}
}