Add utilites to plot Keypoints

[oke-aditya]

🚀 Feature

We added utilities to plot bounding boxes #2556 #2785 and utility to plot semantic segmentation masks #3272 #3330 #3820
I think this is probably the last plotting utility we need.

Motivation

Same as before, we need these to make visualization and post-processing easier.
Also Easily support torchvision models.

Pitch

Use only PIL to draw key points. Also, keep the function signature consistent with previous utilities.

A prototype implementation is here!

@torch.no_grad()
def draw_keypoints(
    image: torch.Tensor,
    keypoints: torch.Tensor,
    colors: Optional[List[Union[str, Tuple[int, int, int]]]] = None,
    radius: Optional[int] = 5,
    connect : Optional[bool] = False,
    plot_ids Optional[bool] = False, 
    font: Optional[str] = None,
    font_size: int = 10
) -> torch.Tensor:

"""
Args:
        image (Tensor): Tensor of shape (C x H x W) and dtype uint8.
        keypoints (Tensor): The ketpoints from Keypoint RCNN
        colors (List[Union[str, Tuple[int, int, int]]]): List containing the colors of keypoints. The colors can
            be represented as `str` or `Tuple[int, int, int]`.
       radius (int): Radius of circles drawn for keypoints.
       connect (bool): If True connects the visible keypoints.
       plot_ids (bool): If True plots the visible keypoints ids on the image.
        font (str): A filename containing a TrueType font. If the file is not found in this filename, the loader may
            also search in other directories, such as the `fonts/` directory on Windows or `/Library/Fonts/`,
            `/System/Library/Fonts/` and `~/Library/Fonts/` on macOS.
        font_size (int): The requested font size in points.
"""

Radius to determine how big the key point should be. This is essential as torchvision models suppose flexible image sizes.
The font is to plot the keypoint id. Sometimes in models like PoseNet, we would like to know which keypoint ids are visible.
Might help in debugging if particular keypoint does not work well.

The output of Keypoint RCNN is as follows

boxes (FloatTensor[N, 4]): the ground-truth boxes in [x1, y1, x2, y2] format,
with values of x between 0 and W and values of y between 0 and H

labels (Int64Tensor[N]): the class label for each ground-truth box

keypoints (FloatTensor[N, K, 3]): the K keypoints location for each of the N instances,
in the format [x, y, visibility], where visibility=0 means that the keypoint is not visible.

I think we should plot only the keypoints which are visible,

Additional context

The boxes can be plotted using draw_bounding_boxes. Users can use both these utils together to plot both boxes and keypoints. I don't think so draw_keypoints should internally call / draw boxes.

Probably we need to give a demo run at it

[oke-aditya]

I think the time has come to work on this. I will illustrate what I have in my mind. I think that boxes can be plotted with utils.draw_bounding_boxes so we won't plot boxes with this utility.

Taken from here

What do you think of this output @datumbox?

This Colab is illustartion of @fmassa 's visualization utils for mask rcnn

Additionally, the key point number would also be plotted if plot_ids is True.

Note:- We would need to implement only with PIL to avoid dependecies !

[datumbox]

@oke-aditya I have not worked much with the keypoint models but it surely looks nice. :)

@fmassa thoughts?

[oke-aditya]

Hey @datumbox
I think I will have bandwidth to work on this (as well as torchvision) from next month (if everything goes well).
Hopefully I will get back with a PR before next month.

[datumbox]

@oke-aditya Apologies for the late response. I'm OOO and don't closely follow the messages. Sounds good, happy to check it out once you are done when I'm back. Thanks for looking into it as always!

[oke-aditya]

No issues @datumbox 😄