Three different datasets are used for the evaluation of EventPoint. In the previous issue 25, we investigate details about three datasets used. In this issue, we investigate metrics used in evaluation.
Precision means, correct matches divide by valid matches(including correct and wrong matches). $\sigma$ is selected as 3, 6, 9 for adjusting the strictness of the matching conditions. $x'$ is the pixel coordinate mapped to the same image as $x$ using gt disparity map provided by DSEC. $N$ is the number of image pairs for evaluation and $M$ is the number of valid matches for each image pair.
IOU Matching Evaluation on N-Caltech101
IOU metric is used for this evaluation,
$IOU = \frac{TP}{TP+FP+FN}$
$TP$, $FP$ and $FN$ denote true positive, false positive and false negative.
Specifically, given two event sequences with a length of L (EventPoint test uses 300ms), the global matching within the object contour is regarded as correct matching, otherwise as wrong matching.
Reprojection Evaluation on HVGA ATIS
HVGA ATIS dataset is used for reprojection evaluation. Specifically, given two 'frame' estimate keypionts and matches, then calculate homography transform. Then use the calculate homography transform to reproject matched points from for example 'frame 2' to 'frame 1' and calculate reprojection precision(unit in pixel).
Three different datasets are used for the evaluation of EventPoint. In the previous issue 25, we investigate details about three datasets used. In this issue, we investigate metrics used in evaluation.
$Precision = \frac{1}{N}\sum{i=1}^{N}\sum{j=1}^{M}\frac{[|x-x'|<\sigma]}{M}$
Precision means, correct matches divide by valid matches(including correct and wrong matches). $\sigma$ is selected as 3, 6, 9 for adjusting the strictness of the matching conditions. $x'$ is the pixel coordinate mapped to the same image as $x$ using gt disparity map provided by DSEC. $N$ is the number of image pairs for evaluation and $M$ is the number of valid matches for each image pair.
IOU metric is used for this evaluation,
$IOU = \frac{TP}{TP+FP+FN}$
$TP$, $FP$ and $FN$ denote true positive, false positive and false negative.
Specifically, given two event sequences with a length of L (EventPoint test uses 300ms), the global matching within the object contour is regarded as correct matching, otherwise as wrong matching.
HVGA ATIS dataset is used for reprojection evaluation. Specifically, given two 'frame' estimate keypionts and matches, then calculate homography transform. Then use the calculate homography transform to reproject matched points from for example 'frame 2' to 'frame 1' and calculate reprojection precision(unit in pixel).