The DMPAC optical flow processing core uses sophisticated methods to compute an accurate and robust measure of the 'estimation confidence' of the computed flow vector and assigns a confidence score to each flow vector based on that.
- Both image and flow field features are used to compute confidence score.
- Following features are used as the basis of confidence measure calculation:
- Winner Cost
- Gradient in the feature descriptor space calculated using hamming
distance
- Flow gradient of horizontal flow
- Flow gradient of vertical flow
- Filtered Cost
- Filtered gradient
- Filtered gradient of horizontal flow
- Filtered gradient of vertical flow
- Fixed 16 independent decision trees are supported for adaptive weight computation.
- Weights are combined linearly and scaled to the supported confidence value range.
- Confidence score value is quantized to 4 bits, 16 levels.