A question about PEViD-UHD dateset about attentionpipeline HOT 4 CLOSED

Thanks for your help. It is very helpful for me to get the right images dataset.

from attentionpipeline.

Hello wyg1997,
I came across similar difficulties when dealing with the PEViD-UHD dataset. The original videos are longer than the annotated segments and if I am not mistaken, there was no easy way how to find where they start.

I extracted frames at 30 fps and then found where each segment starts (manually checking every sequence).

You can try if this works for you as well on one example video "Exchanging_bags_day_indoor_2_original.mp4":

• Convert video to frames at 30 fps:
  ffmpeg -i Exchanging_bags_day_indoor_2_original.mp4 -vf fps=30 frame_%d.jpg

• For this video I found the annotation to fit the range of frames:
  frame_186.jpg to frame_541.jpg

• (Optional) I was working with annotation converted into the format used by LabelImg from the .xgtf format (in this case it would be Exchanging_bags_day_indoor_2_4K.xgtf) which can be also used to manually check the results. The conversion was done using: _side_projects/annotation_conversion/convert_pevid_to_bboxes.py
  If you want to use it, prepare the extracted folder by keeping only files frame_186.jpg to frame_541.jpg and then change paths:
  input_gt_file = "Exchanging_bags_day_indoor_2_4K.xgtf" (downloaded from the website)
  output_folder = "Exchanging_bags_day_indoor_2_original_frames/" (where you extracted frames using ffmpeg)

If this works for you, I can dig around my hard drives and find the ranges for other videos I worked with:

• Exchanging_bags_day_indoor_2
• Exchanging_bags_day_indoor_3
• Exchanging_bags_day_outdoor_4
• Exchanging_bags_day_outdoor_5
• Stealing_day_indoor_2
• Stealing_day_indoor_3
• Stealing_day_indoor_4
• Stealing_day_outdoor_5
• Stealing_day_outdoor_8
• Stealing_day_outdoor_9

I remember this being a bit painful at the start, but when I processed them I was only working with my exported files and didn't need to worry about the original files.

Hope this helps :),
~ Vitek

from attentionpipeline.

^^ This is what you end up eventually getting... (labelImg is quite good to sanity-check the data)

from attentionpipeline.

Using this command:
ffmpeg -i <video file name>.mp4 -vf fps=30 %04d.jpg

from attentionpipeline.