The range of reconstruction score

[usamahjundia]

Hi again, i have another question..

So i wrapped the current reconstruction / prediction network for use in a real-time basis, after carefully making sure it stays true to the original implementation

I think it is also relevant to mention that this is my setup :

• Score from frame t is only based on the reconstruction error of (frame t, frame t-1, ... , frame t - input_length +1), whereby in your original implementation, i think it aggregates reconstruction error of frame t from future and past timesteps too
• the input volume is filled with zeros for missing parts or all zeros for missing detection
• i only used the reconstruction error and not the prediction error, because i need this to run real time (although given enough headstart, i can utilize the prediction error from past frames too)

This might be a silly question, but the reconstruction score is supposed to be within the range of [0, 1], right?

Because testing on another video (different resolution) using the same scalers and the pretrained model for shanghaitech dataset, i observed values bigger than 1 for the reconstruction error, ranging as far as 5!

I just want to know if this is the case or i did something wrong on my preprocessing / postprocessing and score calculation because i wrote custom codes (heavily referring to yours, but just vectorized) for preprocessing (scaling, frame to global and local feats) and postprocessing (flipping) and score calculation

Cheers, thanks in advance

[usamahjundia]

closed because some stupid mistake in preprocessing caused all my global features to be NaN which is then converted to zeros, causing local features to shoot up in value, all fixed now.