Skip to main content

IKeypointSelector

This module selects keypoint given current frame and (optionally) estimated depth, depth_cov, and flow_cov.

The selector also receives an argument numPoint as hint to how many keypoints to select.

info

The numPoint hint may not be followed strictly by the selector. Number of keypoint will fluctuate based on different selection strategy and the input conditions.

warning

Keypoints in this codebase are always arranged in (u, v) format. This means that you need to output the index of keypoints in different coordinate system as pytorch. Use image[kp[..., 1], kp[..., 0]] to read value of image on all u-v coords of keypoints.

Interface​

Module/IKeypointSelector.py
class IKeypointSelector(ABC, ConfigTestableSubclass):
    @abstractmethod
    def select_point(
        self,
        frame: SourceDataFrame,
        numPoint: int,
        depth_map: torch.Tensor | None,
        depth_cov_map: torch.Tensor | None,
        flow_cov_map: torch.Tensor | None,
    ) -> torch.Tensor:
        ...

Methods to Implement​

  • IKeypointSelector.select_point

    Select keypoint for tracking using given frame, (optionally) estimated depth, depth_cov, and flow_cov. Return keypoint as a torch.Tensor with dtype=torch.float and shape (N, 2) where keypoints are arranged in (u, v) format.

Implemenations​

  • class SelectorCompose(IKeypointSelector)

    A higher-order module that combines multiple IKeypointSelector instances with a relative weight. This module distribute keypoint selection to all "sub-selectors" according to the provided weight on every select_point call.

    warning

    The distribution of keypoint selection is achieved by estimating and setting numPoint for each selector included. Therefore, it might not be an accurate split, and the module makes no guarantee to the number of keypoints selected.

  • class RandomSelector(IKeypointSelector)

    Uniformly random select keypoints within the scope of [mask_width : -mask_width]. mask_width is the width of image border that we don't want to select keypoint from.

  • class GradientSelector(IKeypointSelector)

    Select keypoint based on gradient information. Will random select points with sufficiently large local image gradient (calculated by Sobel filter).

  • class SparseGradienSelector(IKeypointSelector)

    Select keypoint based on gradient information. Will random select points with sufficiently large local image gradient (calculated by Sobel filter).

    Ensured sparsity of keypoint by applying non-maximum suppresion (NMS) on image gradient of keypoint candidates.

  • class GridSelector(IKeypointSelector)

    Select keypoint following the grid - strictly uniform across the entire image.

    The requested numPoint will be used to estimate the spacing between keypoints, but the selector may not generate exactly numPoint amount of keypoints.

    info

    This selector is deterministic and is used for internal benchmarking when we want reproducible results.

  • class CovAwareSelector(IKeypointSelector)

    The keypoint selector used by the MAC-VO.

    Selecting keypoints based on estimated depth, depth_cov, and flow_cov. See sect III.B of paper for detail.

  • class CovAwareSelector_NoDepth(IKeypointSelector)

    Selecting keypoints based on estimated flow_cov.

    The main difference with CovAwareSelector is dropping filters related with depth (i.e. max_depth and depth_cov).