Skip to main content

IObservationFilter

This module filters out the observations using some simple logic. It helps for more robust tracking and removes outliers in the system at early stage of the system.

info

This module is not the keypoint filtering strategy introduced in the paper. For the keypoint filtering/selection process, see IKeypointSelector.

Interface​

Module/OutlierFilter.py
class IObservationFilter(ABC, ConfigTestableSubclass):
    def __init__(self, config: SimpleNamespace):
        self.config = config
    
    def set_meta(self, meta: MetaInfo):
        pass

    @abstractmethod
    def filter(self, observations: BatchObservation) -> torch.Tensor:
        ...

Methods to Implement​

  • IObservationFilter.set_meta(self, meta: MetaInfo) -> None

    This method is used to receive meta info (e.g. camera intrinsic, image shape, etc.) on the first frame received by MAC-VO.

    The filter can then initialize some behavior dynamically based on this information.

  • IObservationFilter.filter(self, observations: BatchObservation) -> torch.Tensor

    Given a batch of N observation (BatchObservation), the filter returns a boolean tensor of shape (N,) that

    • sets True for "good" observation
    • sets False for observations to filter away.

Implementations​

  • class FilterCompose(IObservationFilter):

    Compose multiple filters sequentially. Return the logical_and of all sub-filters as final result.

  • class IdentityFilter(IObservationFilter):

    Accept all observations unconditionally.

  • class CovarianceSanityFilter(IObservationFilter):

    Reject all observations with nan or inf value in 3x3 covariance matrix.

  • class SimpleDepthFilter(IObservationFilter):

    Given a min_depth and max_depth value, reject all observations that are out of this range. The max_depth can be set to "auto", in which the maximum depth will be set to baseline * fx to enforce a minimum of 1-pixel disparity.

  • class DepthFilter(IObservationFilter):

    A fancy depth filter that iteratively filter out observations with excessive depth. Prioritize observations with relatively low depth when there are not enough observations in the system.

  • class LikelyFrontOfCamFilter(IObservationFilter):

    Filter out depth where ± 3 std_depth is smaller than 0 (which indicates that the depth estimation is not confidence enough).