search

nguyetn89 / Anomaly_detection_ICCV2019

Anomaly Detection in Video Sequence with Appearance-Motion Correspondence
BSD 2-Clause "Simplified" License
53 stars 15 forks source link

Training details #4

Closed kairikibear closed 4 years ago

kairikibear commented 4 years ago

Hi! Thanks for the hard work. I am currently trying to reimplement this code but what is still ambiguous to me is the following :

  1. Preprocessing of the RGB frames
  2. Preprocessing of the Flow frames

From load_images_and_flows, i saw some processing on the data loaded. It is save to assume that:

  1. For the rgb frames, you divide them by 255, and afterwards you multiply by 2 and then substract 1 to get matrices with value within range of [-1, 1]
  2. For the flow frames, you directly just use the flow outputted by flownet, the only extra processing is adding the magnitude channel

Thanks in advance!

nguyetn89 commented 4 years ago

Hi @kairikibear,

Yes, the normalization of value range is explicitly performed only on RGB frames. In my opinion, the bounding range of optical flow is flexible (not strictly 0-255, 0-1 as RGB pixels), it is thus safer to keep their original values instead of normalizing them.

kairikibear commented 4 years ago

thank you!