I have computed optical flow for an image sequence using H/S method. I have the middlebury as well as the UCL datasets with ground truth flow. My question is, how to compare my results with the ground truth flow ? I have implemented in MATLAB. The UCL datasets have a .flo file and the ground truth image. From this information, how do I get the velocity components to compute the mean absolute error ?
There are two aspects:
1) If you want to compute the optical flow vectors using some method vs the ground truth, here's my advice:
Im assuming these .flo files are in python? If so, they will be in numpy arrays; atleast was in my case. These files are 2D arrays, basically with the x and y coordinate of the flow vector. Once you get your optical flow vector using atleast two images, then use that vector and compare with the vector from the .flo file and do a simple RMSE between the two. And then take average of RMSE across all image pairs in the end to get your final error.
2) If you want to evaluate the velocity components, then my advice is:
Compare the x and y coordinate of optical flow vector from one instant to another instant; each instant is basically two frames, as you need atleast two frames to compute optical flow. Say, for example, compute optical flow 2D array using frame_1 and frame_2 --> gives you the 2D array at instant 1, similarly flow between frame_2 and frame_3 --> gives you 2D array at instant 2. Then the problem becomes as simple as taking the euclidean distance between the x-th component at instant 2 from instant 1 and y-th component at instant 2 from instant 1. Finally, you will have a speed of certain pixel units/frame. You can convert it to pixel units/second (if you assume 1s = 30 frames).
Hope it helps!
Tahmid Chowdhury I am currently confined to the first aspect. But the problem is I do not know how to extract the components from the .flo files. Also another thing is how to create a ground truth image from the flow? I am attaching a specific .flo file with the corresponding ground truth image.
You can follow this link https://github.com/georgegach/flowiz to convert .flo files to color images. It's pretty simple.
Tahmid Chowdhury How might you calculate the velocity of an object from a ".flo" output (I used deep learning approach to generate optical flow. Output is a ****.flo file which in turn is converted to a color image for visualization)?
- Badges
- Science topic