def run(video_id, video): ''' Main function for running all feature extraction steps. TODO: move this to "__main__" ARGS: video_id: unique video identifier video: path to video file (mp4 format) RETURNS: dataframe with frame-by-frame analysis ''' ## Go through every frame in the video and extract features cap = cv2.VideoCapture(video) frame = None ## Output dataframes video_df = pd.DataFrame() img_quality_df = pd.DataFrame() num_frames = cap.get(cv2.cv.CV_CAP_PROP_FRAME_COUNT) #num_frames = 3 debug = 0 if debug == 0: i_frame = 0 while i_frame < num_frames: ## Progress logging if i_frame % 100 == 0: print i_frame if frame is not None: prev_frame = frame.copy() else: prev_frame = None ret,frame = cap.read() ## If there are no more frames, break out of loop if ret == False: break frame_series = pd.DataFrame(index=[0]) frame_series['video_id'] = video_id frame_series['frame_number'] = i_frame frame_series['time'] = cap.get(cv2.cv.CV_CAP_PROP_POS_MSEC) / 1000. ##======================= ## Image quality metrics ##======================= ## Blur blur_df = SceneAnalysis.get_blur(frame) ## Color Spectrum color_df = SceneAnalysis.get_hsv_hists(frame) img_quality_df = img_quality_df.append(pd.concat([frame_series, blur_df, color_df], axis=1), ignore_index=True) #img_quality_df = img_quality_df.append(frame_series) ##======================= ## Video/motion metrics ##======================= ## Optical Flow if prev_frame is not None: flow_df = VideoUtilities.optical_flow_on_frame(frame, prev_frame) else: flow_df = None ## Merge results into single dataframe #video_df = video_df.append(pd.concat([frame_series, blur_df, color_df, flow_df], axis=1, ignore_index=True)) #video_df = video_df.append(frame_series) #cat['video_id'] = video_id #cat = pd.concat([blur_df, color_df, flow_df], axis=1) if flow_df is not None: flow_series = flow_df.copy() ##NOTE: the below code is duplicated above - bad style flow_series['video_id'] = video_id flow_series['frame_number'] = i_frame #cap.get(cv2.cv.CV_CAP_PROP_POS_FRAMES-1) flow_series['time'] = cap.get(cv2.cv.CV_CAP_PROP_POS_MSEC) / 1000. #cat = pd.concat([frame_series, cat], axis=1) video_df = video_df.append(flow_series, ignore_index=True) i_frame += 1 else: img_quality_df = pd.read_pickle(video_id + '.img_quality.pkl') ## After completing the frame-by-frame analysis, run video metrics ## scene changes scene_change_df = VideoAnalysis.detect_cut(video, img_quality_df['time']) img_quality_df['is_scene_transition'] = scene_change_df['is_scene_transition'].copy() ## Pickle model and save it to S3 or local directory img_quality_df.to_pickle(video_id + '.img_quality.pkl') video_df.to_pickle(video_id + '.flow.pkl') return img_quality_df, video_df
def run(video_id, video): ''' Main function for running all feature extraction steps. TODO: move this to "__main__" ARGS: video_id: unique video identifier video: path to video file (mp4 format) RETURNS: dataframe with frame-by-frame analysis ''' ## Go through every frame in the video and extract features cap = cv2.VideoCapture(video) frame = None ## Output dataframes video_df = pd.DataFrame() img_quality_df = pd.DataFrame() num_frames = cap.get(cv2.cv.CV_CAP_PROP_FRAME_COUNT) #num_frames = 3 debug = 0 if debug == 0: i_frame = 0 while i_frame < num_frames: ## Progress logging if i_frame % 100 == 0: print i_frame if frame is not None: prev_frame = frame.copy() else: prev_frame = None ret, frame = cap.read() ## If there are no more frames, break out of loop if ret == False: break frame_series = pd.DataFrame(index=[0]) frame_series['video_id'] = video_id frame_series['frame_number'] = i_frame frame_series['time'] = cap.get( cv2.cv.CV_CAP_PROP_POS_MSEC) / 1000. ##======================= ## Image quality metrics ##======================= ## Blur blur_df = SceneAnalysis.get_blur(frame) ## Color Spectrum color_df = SceneAnalysis.get_hsv_hists(frame) img_quality_df = img_quality_df.append(pd.concat( [frame_series, blur_df, color_df], axis=1), ignore_index=True) #img_quality_df = img_quality_df.append(frame_series) ##======================= ## Video/motion metrics ##======================= ## Optical Flow if prev_frame is not None: flow_df = VideoUtilities.optical_flow_on_frame( frame, prev_frame) else: flow_df = None ## Merge results into single dataframe #video_df = video_df.append(pd.concat([frame_series, blur_df, color_df, flow_df], axis=1, ignore_index=True)) #video_df = video_df.append(frame_series) #cat['video_id'] = video_id #cat = pd.concat([blur_df, color_df, flow_df], axis=1) if flow_df is not None: flow_series = flow_df.copy() ##NOTE: the below code is duplicated above - bad style flow_series['video_id'] = video_id flow_series[ 'frame_number'] = i_frame #cap.get(cv2.cv.CV_CAP_PROP_POS_FRAMES-1) flow_series['time'] = cap.get( cv2.cv.CV_CAP_PROP_POS_MSEC) / 1000. #cat = pd.concat([frame_series, cat], axis=1) video_df = video_df.append(flow_series, ignore_index=True) i_frame += 1 else: img_quality_df = pd.read_pickle(video_id + '.img_quality.pkl') ## After completing the frame-by-frame analysis, run video metrics ## scene changes scene_change_df = VideoAnalysis.detect_cut(video, img_quality_df['time']) img_quality_df['is_scene_transition'] = scene_change_df[ 'is_scene_transition'].copy() ## Pickle model and save it to S3 or local directory img_quality_df.to_pickle(video_id + '.img_quality.pkl') video_df.to_pickle(video_id + '.flow.pkl') return img_quality_df, video_df
flow_series = flow_df.copy() ##NOTE: the below code is duplicated above - bad style flow_series['video_id'] = video_id flow_series['frame_number'] = i_frame #cap.get(cv2.cv.CV_CAP_PROP_POS_FRAMES-1) flow_series['time'] = cap.get(cv2.cv.CV_CAP_PROP_POS_MSEC) / 1000. #cat = pd.concat([frame_series, cat], axis=1) video_df = video_df.append(flow_series, ignore_index=True) i_frame += 1 else: img_quality_df = pd.read_pickle(video_id + '.img_quality.pkl') ## After completing the frame-by-frame analysis, run video metrics ## scene changes scene_change_df = VideoAnalysis.detect_cut(video, img_quality_df['time']) img_quality_df['is_scene_transition'] = scene_change_df['is_scene_transition'].copy() ## Pickle model and save it to S3 or local directory img_quality_df.to_pickle(video_id + '.img_quality.pkl') video_df.to_pickle(video_id + '.flow.pkl') return img_quality_df, video_df if __name__ == "__main__": video = "../media/CKeLfaOl0Qk.mp4" video_df = pd.read_pickle('CKeLfaOl0Qk.img_quality.pkl') video_series = video_df['time'] scene_change_df = VideoAnalysis.detect_cut(video, video_series) #print scene_change_df[scene_change_df['is_scene_transition'] == 1] video_df['is_scene_transition'] = scene_change_df['is_scene_transition'].copy() print video_df[video_df['is_scene_transition'] == 1]
cv2.cv.CV_CAP_PROP_POS_MSEC) / 1000. #cat = pd.concat([frame_series, cat], axis=1) video_df = video_df.append(flow_series, ignore_index=True) i_frame += 1 else: img_quality_df = pd.read_pickle(video_id + '.img_quality.pkl') ## After completing the frame-by-frame analysis, run video metrics ## scene changes scene_change_df = VideoAnalysis.detect_cut(video, img_quality_df['time']) img_quality_df['is_scene_transition'] = scene_change_df[ 'is_scene_transition'].copy() ## Pickle model and save it to S3 or local directory img_quality_df.to_pickle(video_id + '.img_quality.pkl') video_df.to_pickle(video_id + '.flow.pkl') return img_quality_df, video_df if __name__ == "__main__": video = "../media/CKeLfaOl0Qk.mp4" video_df = pd.read_pickle('CKeLfaOl0Qk.img_quality.pkl') video_series = video_df['time'] scene_change_df = VideoAnalysis.detect_cut(video, video_series) #print scene_change_df[scene_change_df['is_scene_transition'] == 1] video_df['is_scene_transition'] = scene_change_df[ 'is_scene_transition'].copy() print video_df[video_df['is_scene_transition'] == 1]