Can mmpose support the rotated bbox in mmrotate?

[qwert31639]

請問是不是有機會支援mmrotate中的旋轉bbox呢？
數據集可以利用帶有mask以及skeleton的數據，生成旋轉角度。
下面是我轉換bbox的代碼

import numpy as np
from typing import List, Optional, Sequence, Tuple, Union
import cv2

def make_bbox(segpoint: np.ndarray,
              headpt: np.ndarray,
              footpt: np.ndarray) -> Tuple:
    '''
    input 
        segpoint: shape(-1,2)
        We can use the mask annotations in the coco dataset, 
        or use the bounding box's four boundary points as a replacement.
        headpoint: shape(2,)
        This represents the center of the key points for the human head.
        like eyes(mean axis), ears(mean axis), nose
        footpoint: shape(2,) 
        This represents the center of the key points for the human foot.
        like ankles(mean axis), knees(mean axis)
    output 
        rotate_bbox: tuple(centerx, centery, height, width, angle(degree))
        poly_bbox: np.ndarray shape=(4, 2)
    '''
    h_,f_ = headpt, footpt
    vecs = (headpt-footpt)[None,:].astype(np.float32)
    points = segpoint[None, ...]
    mags = np.sqrt((vecs*vecs).sum(axis=-1, keepdims=True))
    mags[mags==0] = 1
    vecs /= mags
    perps = vecs[:,::-1].copy().astype(np.float32)
    perps[:,0] *= -1
    sidevecs = points - footpt
    scores = (perps[:,None,:] * sidevecs).sum(axis=-1)
    maxes = scores.max(axis=-1)
    mins = scores.min(axis=-1)
    maxes[maxes<0] = 0
    mins[mins>0] = 0
    widths = (maxes-mins)[:,None]
    cxy = footpt + perps*((maxes+mins)*0.5)[:,None]
    scores = (vecs[:,None,:] * sidevecs).sum(axis=-1)
    maxes = scores.max(axis=-1)
    mins = scores.min(axis=-1)
    maxes[maxes<0] = 0
    mins[mins>0] = 0
    heights = (maxes - mins)[:,None]
    cxy += (0.5*(maxes+mins))[:,None]*vecs
    angles = np.arctan2(perps[...,1], perps[...,0])
    angles = (angles/np.pi*180)+ 90
    rotate_bbox = ((cxy[0][0], cxy[0][1]), (heights[0][0], widths[0][0]), angles[0])
    poly_bbox   = cv2.boxPoints(rotate_bbox)
    return rotate_bbox
info = np.load('./segpoint.npy',allow_pickle=True)
poly = make_bbox(info[0],info[1],info[2])

上面這張圖是轉換的結果，
白色線段：軸心線
紅色線段：cv2.minAreaRect的結果
黃色線段：上面代碼轉換的bbox

[ly015]

Hi, is the purpose to estimate the roll angle of the bounding box from the skeleton? I think it would be a useful tool. Would you be willing to submit a pull request to integrate this function into MMPose?