Added optional Keypoint OKS sigmas argument

[sampepose]

This change allows adding custom keypoint sigmas to compute keypoint OKS. A real-world use case is computing metrics on extended keypoint models, such as 19 keypoints (adding head-top and chin to the standard COCO 17 keypoints).

[rbgirshick]

I'd like to merge this simple PR to improve support for evaluating custom keypoint datasets that have different keypoint definitions. @pdollar, @tylin, @matteorr any objections?

[matteorr]

@rbgirshick - It seems fine with me. However, wouldn't it be better to use the params class instead of passing it like that? I.e., take a look at how I did it in the cocoanalyze repo.

coco_eval = COCOeval(coco_gt, coco_dt, 'keypoints')
coco_eval.params.kpt_oks_sigmas = np.array([.26, .25, .25, .35, .35, .79, .79, .72, .72, .62,.62, 1.07, 1.07, .87, .87, .89, .89])/10.0

This way we can keep the init iou_type agnostic.

[pdollar]

Sounds good to me. Will let @tylin comment on specifics for the python version of the API.

[sampepose]

Do you want any more changes, @tylin / @matteorr?

[matteorr]

Everything looks fine to me! Thanks for doing this.