Why the annotated keypoints are sometimes very small or huge?

[futakw]

Objectron/objectron/dataset/graphics.py

Lines 34 to 40 in c06a651

	# The keypoints are [x, y, d] where `x` and `y` are normalized (`uv`-system)\
	# and `d` is the metric distance from the center of the camera. Convert them
	# keypoint's `xy` value to pixel.
	keypoints = [
	np.multiply(keypoint, np.asarray([w, h, 1.], np.float32)).astype(int)
	for keypoint in keypoints
	]

As described here, I think the "keypoints.point_2d.x" or "keypoints.point_2d.y" should be in the range of 0 to 1 if x/y are inside the image.
However, I observe that sometimes those are extremely small or huge.

For example,
"~/bike/batch-11/5/annotation.pbdata"
has

keypoints {                                                                                                                                                                                               
    id: 2                                                                                                                                                                                                   
    point_3d {                                                                                                                                                                                              
      x: 1.2235139608383179                                                                                                                                                                                 
      y: 1.1318135261535645                                                                                                                                                                                 
      z: 0.07509636878967285                                                                                                                                                                                
    }                                                                                                                                                                                                       
    point_2d {                                                                                                                                                                                              
      x: -15.717081069946289                                                                                                                                                                                
      y: -12.666918754577637                                                                                                                                                                                
      depth: -0.07509636878967285                                                                                                                                                                           
    }                                                                                                                                                                                                       
  }  

with extremely small x and y.

[ahmadyan]

You can discard the point when the depth is negative (i.e. the point is behind the camera).