fusing two-stream predictions about coclr HOT 1 CLOSED

Hi,

For each video, first we get the predicted probability (101-D probability vector, sum up to 1) from both streams (here we get 2x101 probability vector).

Then we average these two probability vectors for each video (2x101 --> 1x101).

Then we get the top1/top5 predicted class for each video, then calculate accuracy for the entire testing set.

Example code is here: https://github.com/TengdaHan/CoCLR/blob/main/eval/merge_2stream_prob.py

To use this file, you have to save the output results following the instructions. Of course, you can write one script for yourself to do the average if it's more convenient.

This method was initially used in Two-Stream Network (https://arxiv.org/abs/1406.2199).

from coclr.