Given, two frames from a real world scenario such as a surveillance video of a street where there are pedestrians moving. I need to compare the performance of different optical flow algorithms ( Lucas-Kanade, TV-L1 [Zach et al.], Brox et al.) over these datas.
Are there any standard techniques for the evaluation of the flow vectors obtained from each when applied to this data?
(There is the middlebury evaluation for this purpose but it does not contain any real world data)
If there are no existing techniques, then, how to create a ground truth manually for such real world data?
Hi Abhilash,
One way to test the accuracy of your estimation is to compare the original frame and the motion compensated frame. Say you have two frames F_t and F_{t+1} and you have computed forward (or backward) motion flow between these two frames. Now, you warp one frame using the motion vectors and try to reconstruct the other frame. This is simplest motion compensation algorithm. Say from F_t, you estimate the motion compensated frame M_{t+1}. Now compare M_{t+1} and F_{t+1}. There are several algorithms to compare them. You can start with PSNR, SSIM etc. Pick the optical flow algorithm that will give you the best estimation.
You may have to check with different video data sets to have a definite answer and it is also likely that for different datasets, different algorithms may give you the best predictions. So try to perform the experiment on several videos and decide the winner.
Take a look at this article:
http://scribblethink.org/Work/V/FLOWEVAL.pdf
From what I observed, end point error seems to be the method that is most popular currently (at least for dense optical flow).
- Badges
- Science topic
- Similar topics
- Vectorization