questions about the performance

[newdpdeeplearner]

excuse me, i've read your paper which is impressive. so i run your code on my own machine. While the result of the proposed model is reproducible, the results of the 'Chance' baseline are much worse as comparted to those reported in your paper. Specifically, the random temporal boundaries were generated through 'start = tf.random_uniform([], maxval=length, dtype=tf.int32), end = tf.random_uniform([], minval=start, maxval=length+1, dtype=tf.int32)' where 'length' is the mapped video_length of 'len_r' in your code. However, the performance are as follows:

Saving dict for global step 200: IoU/0.1 = 0.42229038, IoU/0.2 = 0.32719836, IoU/0.3 = 0.23721881, IoU/0.4 = 0.17177914, IoU/0.5 = 0.111451946, IoU/0.6 = 0.06543967, IoU/0.7 = 0.03578732, IoU/0.8 = 0.020449897, IoU/0.9 = 0.0040899795, IoU/mean = 0.047443762

i wonder whether my implementation is right or not. Thank you

[fengyang0317]

Could you first sample length uniformly and then sample the start position?

[newdpdeeplearner]

i'm sorry, what does 'sample length uniformly' mean ? i mean, isn't the length of each reference video is a fixed number? or you mean i should calculate the average length of all the moments in the dataset, and then sample the start position and determine the video segment with the average length?

[fengyang0317]

If the length of the reference is L, sample the target length in the range [1, L] uniformly.

[newdpdeeplearner]

thank you very much!