MoonBlvd / tad-IROS2019

Code of the Unsupervised Traffic Accident Detection paper in Pytorch.
MIT License
164 stars 39 forks source link

A3D Dataset - There are videos that are not used. #11

Open seonho opened 5 years ago

seonho commented 5 years ago

There are videos (A3D_urls.txt) that are not used in A3D_labels.pkl.

What I've done is as follows.

from joblib import load

data = load('datasets/A3D_labels.pkl')

used = set([value['video_name'] for _, value in data.items()])

prefix = len('https://www.youtube.com/watch?v=')

with open('datasets/A3D_urls.txt') as f:
    lines = f.read().split()
    total = [line[prefix:] for line in lines]

unused = set(total) - set(used)
print(len(total), len(used), len(unused))

Then, I got in total 204 videos, 80 are used and 124 are not.

MoonBlvd commented 5 years ago

Hi @seonho, yes we have only labeled part of those videos because there are too many of them. In facting we are annotating the rest now and will release them with our next paper which will come out soon. So far we have labelled 1500 video clips from the 80 long videos you saw, and the label is provided in this repo. We will update this repo about the new data/paper later this year. Thanks!