Draw results on images

Every drawing function lives in pf.annotate and takes the same two things: the image, and the results to draw. Each one draws one kind of mark (boxes, labels, masks...), so you build a picture by calling them one after the other.

Python
import pixelflow as pf
from mayaku import from_pretrained

predictor = from_pretrained("mayaku-n-seg")
result = predictor("street.jpg")
detections = pf.from_mayaku(result, labels=predictor.class_names)
detections = detections.filter_by_confidence(0.5)

image = pf.read_image("street.jpg")
image = pf.annotate.mask(image, detections)
image = pf.annotate.box(image, detections)
image = pf.annotate.label(image, detections)
pf.save_image("street-annotated.jpg", image)

Masks, boxes and labels drawn on the people, motorcycles and an umbrella in a street photo

Order matters

Each call paints over what is already there. Draw the large, see-through marks first (masks, filled boxes, zones) and the small, solid ones last (boxes, then labels), so the labels stay readable.

Drawing changes the image you pass in

The functions draw directly on the array you give them and return that same array. This avoids a copy of every frame in a video, but it also means the original is gone. Pass a copy when you still need it:

Python
original = pf.read_image("street.jpg")
annotated = pf.annotate.box(original.copy(), detections)

Images are RGB, as pf.read_image and pf.VideoReader return them. An image read with cv2.imread is BGR: the drawing works, but every color comes out with red and blue swapped.

Sizes follow the image

Line widths, font sizes, blur strength and the other sizes are worked out from the image size, so the same code gives readable marks on a 640×480 frame and on a 4K photo. On the 1920×1280 photo above, lines are 4 pixels wide. Every function takes the size as an argument (thickness, font_scale, radius...) when you want to set it yourself.

Colors follow the class

Each detection takes its color from its class_id, so every person has the same color, in every function and in every frame of a video. With the default palette of 20 colors, class n gets color n % 20. A detection without a class_id gets the first color.

Pass colors to use your own list of RGB colors, with the same rule:

Python
image = pf.annotate.box(image, detections, colors=pf.colors.PASTEL_PALETTE)
image = pf.annotate.box(image, detections, colors=[(255, 0, 0)])  # everything red

pf.colors has three palettes: DEFAULT_PALETTE, VIBRANT_PALETTE and PASTEL_PALETTE.

All drawing functions

Function Draws Needs
box the outline of each box boxes
filled_box see-through filled boxes boxes
label text next to each box boxes
mask see-through masks masks
polygon the outline of each object outlines
oval an ellipse under each object boxes
anchors dots at the corners, center or edges of each box boxes
keypoint a dot on each keypoint keypoints
keypoint_skeleton lines between keypoints named keypoints
classification the top classes in a corner classifications
blur each box blurred boxes
pixelate each box pixelated boxes
zones zones and their counts Zones
crossings lines and their in/out counts Crossings
grid_overlay a grid inside each box boxes
fps_counter frames per second a video loop