Filtering Depth Values for ZED SDK VIO (WIP) #54
Description
🎯 Objective
Using the ZED SDK's python API, try to find a way to infer object bounding boxes's depth values to deduce that the object's most accurate depth values is. Program an extension to our model inference code to reflect that more accurately.
📖 Description
Program a node that adds to the current inference code that would deduce and filter out the depth map such that our object position estimation in the 3D space is more accurate. Verify and compare this performance with the origin scheme and document those changes.
✅ Acceptance Criteria
- New Object Position is more accurate and drifts less
- Deduce a test to see how much the projected object locations are drifting from the ground truth (ground truth pending from the simulation)
- Document those changes
🛠️ Implementation
Approach:: Use some sort of OpenCV algorithm to clean out the bounding box? You could also apply a Gaussian kernel on the bounding box (after overlapping with a depth map) to smooth out the background. Alternative solutions might exist because. I'm not specialized in this field.
Dependencies: [Dependencies]
🔗 Links
Related: [#issue], [#issue]
Blocks: [#issue]
Blocked by: [#issue]