TDA4VE-Q1: DOF Small Pixel Tracking

Hello Laron, 

Can you please share details for the use-case scenario? details such as- 

1) Background intensity

2) Object vs background contrast

3) Object size

4) Motion magnitude

5) Number of objects

How close is the synthetic test case to the real use-case scenario? 

Thanks and Regards

Hrushikesh

Thanks for the reply. Here are the details you requested:

1. Background intensity - 128 (gray)
2. Object vs background contrast - Full contrast (gray object on black background)
3. Object size - Square of 9x9px or smaller
4. Motion magnitude - 1 to 7px per frame
5. Number of objects - Single object

This synthetic test case should be much easier to detect than a real world case which includes the presence of noise, less contrast, and potentially smaller areas. When the square shrinks below a certain size (around 7px), the motion flow no longer registers, even with changes to the median filter and motionSmoothnessFactor settings.

Hi Laron, 

Did you mean: 

1. Background intensity - 0
2. Object vs background contrast - Full contrast (gray object with pixel value 128 on black background)

In general noise, contras, object size, structure of the scene (distance of object and its local background from the camera) and ego and relative motion all do influence optical flow estimation and accuracy.  Are you facing any specific challenge in meeting quality requirements?  Will you be able to share the use-case test image pair with us?

Hrushikesh

Correct, my mistake - background intensity was 0 (black) and the object was gray (128).

At this point we are mainly evaluating the DMPAC DOF for feasibility against our use case. When certain motion flow output with actual data didn't make sense, we looked at a synthetic example in an attempt to isolate and gain more visibility. If the DOF is unable to detect a small moving rectangle against a black background, then I am confident it won't work in our intended use case either. Is that something that TI would expect to be detectable with the DOF implementation? If we are able to decide definitively then we could possibly move on to more realistic cases.

Please see also the image in TDA4VE-Q1: DOF small pixel tracking - Processors forum - Processors - TI E2E support forums which is a good reference point:

If I understood correctly, you have tested the DOF in real test scenario and would like to improve the accuracy beyond what you have observed? 

Correct.

Any other thoughts on this? It doesn't seem like the DOF will be able to support our use case if it is unable to track the small rectangles as in the synthetic input. Thanks for any insights you might be able to provide.