diff options
| -rw-r--r-- | youtube_dl/extractor/__init__.py | 1 | ||||
| -rw-r--r-- | youtube_dl/extractor/veoh.py | 47 | 
2 files changed, 48 insertions, 0 deletions
| diff --git a/youtube_dl/extractor/__init__.py b/youtube_dl/extractor/__init__.py index f668f0f4a..9aaa7ca9b 100644 --- a/youtube_dl/extractor/__init__.py +++ b/youtube_dl/extractor/__init__.py @@ -58,6 +58,7 @@ from .tumblr import TumblrIE  from .tutv import TutvIE  from .ustream import UstreamIE  from .vbox7 import Vbox7IE +from .veoh import VeohIE  from .vevo import VevoIE  from .vimeo import VimeoIE  from .vine import VineIE diff --git a/youtube_dl/extractor/veoh.py b/youtube_dl/extractor/veoh.py new file mode 100644 index 000000000..c1f8271f1 --- /dev/null +++ b/youtube_dl/extractor/veoh.py @@ -0,0 +1,47 @@ +import re +import json + +from .common import InfoExtractor +from ..utils import ( +    determine_ext, +) + +class VeohIE(InfoExtractor): +    _VALID_URL = r'http://www\.veoh\.com/watch/v(?P<id>\d*)' + +    _TEST = { +        u'url': u'http://www.veoh.com/watch/v56314296nk7Zdmz3', +        u'file': u'56314296.mp4', +        u'md5': u'620e68e6a3cff80086df3348426c9ca3', +        u'info_dict': { +            u'title': u'Straight Backs Are Stronger', +            u'uploader': u'LUMOback', +            u'description': u'At LUMOback, we believe straight backs are stronger.  The LUMOback Posture & Movement Sensor:  It gently vibrates when you slouch, inspiring improved posture and mobility.  Use the app to track your data and improve your posture over time. ', +        } +    } + +    def _real_extract(self, url): +        mobj = re.match(self._VALID_URL, url) +        video_id = mobj.group('id') +        webpage = self._download_webpage(url, video_id) + +        m_youtube = re.search(r'http://www\.youtube\.com/v/(.*?)(\&|")', webpage) +        if m_youtube is not None: +            youtube_id = m_youtube.group(1) +            self.to_screen(u'%s: detected Youtube video.' % video_id) +            return self.url_result(youtube_id, 'Youtube') + +        self.report_extraction(video_id) +        info = self._search_regex(r'videoDetailsJSON = \'({.*?})\';', webpage, 'info') +        info = json.loads(info) +        video_url =  info.get('fullPreviewHashHighPath') or info.get('fullPreviewHashLowPath') +        self.to_screen(video_url) +        return {'id': info['videoId'],  +                'title': info['title'], +                'ext': determine_ext(video_url), +                'url': video_url, +                'uploader': info['username'], +                'thumbnail': info.get('highResImage') or info.get('medResImage'), +                'description': info['description'], +                'view_count': info['views'], +                } | 
