correct typo (per-keypont -> per-keypoint)

dataset/keypoints-eval.htm

@@ -14,7 +14,7 @@ <h2>1.2. Object Keypoint Similarity</h2>
     <b>OKS = Σ<sub>i</sub>[exp(-d<sub>i</sub><sup>2</sup>/2s<sup>2</sup>&kappa;<sub>i</sub><sup>2</sup>)&delta;(v<sub>i</sub>>0)] / Σ<sub>i</sub>[&delta;(v<sub>i</sub>>0)]</b>
   </div>
 </div>
-<p>The d<sub>i</sub> are the Euclidean distances between each corresponding ground truth and detected keypoint and the v<sub>i</sub> are the visibility flags of the ground truth (the detector's predicted v<sub>i</sub> are not used). To compute OKS, we pass the d<sub>i</sub> through an unnormalized Guassian with standard deviation s&kappa;<sub>i</sub>, where s is the object scale and &kappa;<sub>i</sub> is a per-keypont constant that controls falloff. For each keypoint this yields a keypoint <i>similarity</i> that ranges between 0 and 1. These similarities are averaged over all labeled keypoints (keypoints for which v<sub>i</sub>>0). Predicted keypoints that are not labeled (v<sub>i</sub>=0) do not affect the OKS. Perfect predictions will have OKS=1 and predictions for which all keypoints are off by more than a few standard deviations s&kappa;<sub>i</sub> will have OKS~0. The OKS is analogous to the IoU. Given the OKS, we can compute AP and AR just as the IoU allows us to compute these metrics for box/segment detection.</p>
+<p>The d<sub>i</sub> are the Euclidean distances between each corresponding ground truth and detected keypoint and the v<sub>i</sub> are the visibility flags of the ground truth (the detector's predicted v<sub>i</sub> are not used). To compute OKS, we pass the d<sub>i</sub> through an unnormalized Guassian with standard deviation s&kappa;<sub>i</sub>, where s is the object scale and &kappa;<sub>i</sub> is a per-keypoint constant that controls falloff. For each keypoint this yields a keypoint <i>similarity</i> that ranges between 0 and 1. These similarities are averaged over all labeled keypoints (keypoints for which v<sub>i</sub>>0). Predicted keypoints that are not labeled (v<sub>i</sub>=0) do not affect the OKS. Perfect predictions will have OKS=1 and predictions for which all keypoints are off by more than a few standard deviations s&kappa;<sub>i</sub> will have OKS~0. The OKS is analogous to the IoU. Given the OKS, we can compute AP and AR just as the IoU allows us to compute these metrics for box/segment detection.</p>
 
 <h2>1.3. Tuning OKS</h2>
 <p>We tune the &kappa;<sub>i</sub> such that the OKS is a perceptually meaningful and easy to interpret similarity measure. First, using 5000 redundantly annotated images in val, for each keypoint type i we measured the per-keypoint standard deviation &sigma;<sub>i</sub> with respect to object scale s. That is we compute <b>&sigma;<sub>i</sub><sup>2</sup>=E[d<sub>i</sub><sup>2</sup>/s<sup>2</sup>]</b>. &sigma;<sub>i</sub> varies substantially for different keypoints: keypoints on a person's body (shoulders, knees, hips, etc.) tend to have a &sigma; much larger than on a person's head (eyes, nose, ears).</p>