Skip to content

Exploring Polylines – FiftyOne Tips and Tricks – September 22nd, 2023

Welcome to our weekly FiftyOne tips and tricks blog where we cover interesting workflows and features of FiftyOne! This week we are taking a look at polylines. We aim to cover the basics of creating a polyline and how they can be utilized to create specialized labels.

Wait, what’s FiftyOne?

FiftyOne is an open source machine learning toolset that enables data science teams to improve the performance of their computer vision models by helping them curate high quality datasets, evaluate models, find mistakes, visualize embeddings, and get to production faster.

Ok, let’s dive into this week’s tips and tricks! Also feel free to follow along in our notebook or on YouTube!

What is a Polyline?

In computer vision, a polyline is a sequence of connected line segments used to represent and approximate the shape of an object or region within an image. Polylines are commonly employed in computer vision tasks such as image annotation, object tracking, and shape analysis. Polylines can be used to represent the trajectory of an object or map and route representations. Polylines can also be filled to represent different polygons’  regions of interest. With FiftyOnes implementation of polylines, labels can be as creative as you want!

Creating Polylines in Your Samples

import fiftyone as fo
import fiftyone.zoo as foz

dataset = foz.load_zoo_dataset(

session = fo.launch_app(dataset)

Using the app, I am going to tag a non-busy picture for use in the example. To do so, simply select the image, click on the tag image and add “polylines” to its sample tags

Let’s walk through some examples and how they can be used in various use cases. We grab our sample to draw on it next.

example_view = dataset.match_tags("polylines")
sample = example_view.first()

Classic Polylines

The most basic polyline is a line defined by its points. It could be pointing in the direction of a trajectory, or simply be a label for the divide between two regions of interest. To define a Polyline label, provide a list of points in (x,y) bounded by (0,1) for x and y. We also pass the argument closed=False to denote that we are not closing the polyline, and filled=False to likewise express that we don’t want the polyline to be filled.

# A simple polyline
polyline1 = fo.Polyline(
   points=[[(0.3, 0.3), (0.7, 0.3), (0.7, 0.3)]],

Classic Polygon with Polylines

Similar to the last example, we can define polygons with a set of points and close and fill the shape.  Just like with any other label, we can add custom attributes to the polyline as well. In this case I will create a triangle and specify it as a right triangle. 

💡 To close a shape, only the vertices are required and there is no need to specify the first vertex again.

We can add these to our sample using Polylines, a group of polylines, and save our view to look at our new polylines. If it makes it easier to see, you can turn off seeing the ground_truth or prediction labels by unchecking their box under labels in the app.

# A closed, filled polygon with a label
polyline2 = fo.Polyline(
   points=[[(0.1, 0.1), (0.3, 0.1), (0.3, 0.3)]],
   kind="right",  # custom attribute

sample["polylines"] = fo.Polylines(polylines=[polyline1, polyline2])
session.view = example_view


Cuboids, or 3D bounding boxes, are essential in computer vision for a range of applications. They serve to detect and localize objects in a three-dimensional environment, making them valuable in autonomous driving, object tracking, augmented reality, robotics, depth sensing, gesture recognition, human pose estimation, indoor navigation, industrial automation, and virtual reality. Cuboids help model the 3D extent of objects, enabling accurate perception, tracking, and interaction with the three-dimensional world in various domains.

Today, we will be looking at how we can use polylines to create cuboids on our 2D image. FiftyOne does support 3D boxes for point clouds, but that will be covered another day.

 We define a function that will create a random cuboid label for our sample. Cuboids are defined with 8 points in the following sequence:

   7------------ 6
  /|   	     /|
 / |  	    /  |
3--------  2     |
|  4-----  |---- 5
| /  	  |   /
|/   	  | /
0--------  1

Using the fo.Polyline.from_cuboid() method, we are able to easily create a cuboid once we have our 8 points.

import numpy as np

def random_cuboid():
   x0, y0 = [0, 0.2] + 0.8 * np.random.rand(2)
   dx, dy = (min(0.8 - x0, y0 - 0.2)) * np.random.rand(2)
   x1, y1 = x0 + dx, y0 - dy
   w, h = (min(1 - x1, y1)) * np.random.rand(2)
   front = [(x0, y0), (x0 + w, y0), (x0 + w, y0 - h), (x0, y0 - h)]
   back = [(x1, y1), (x1 + w, y1), (x1 + w, y1 - h), (x1, y1 - h)]
   return fo.Polyline.from_cuboid(front + back, label="cuboid")

sample["polyline"] = random_cuboid()
session.view = example_view

Here is what a finished result could look like:

Rotated Bounding Box

Rotated bounding boxes are essential for tasks like object detection, localization, and tracking, particularly when objects are not aligned with the standard horizontal and vertical axes. They’re valuable in applications such as text detection in images, scene text recognition, and irregular object localization. These bounding boxes represent objects’ orientations more accurately, allowing for precise positioning and analysis in situations where traditional axis-aligned boxes wouldn’t suffice. To use a rotating bounding box in FiftyOne, we can use fo.Polyline.from_rotated_box(), where we provide the center of the box in xc,yc as well as the width, height, and the angle at which the box in rotated.

def random_rotated_box():
   xc, yc = 0.2 + 0.6 * np.random.rand(2)
   w, h = 1.5 * (min(xc, yc, 1 - xc, 1 - yc)) * np.random.rand(2)
   theta = 2 * np.pi * np.random.rand()

   return fo.Polyline.from_rotated_box(xc, yc, w, h, theta, label="box")

sample["polyline"] = random_rotated_box()
session.view = example_view

A finished result I ran was this:


In conclusion, polylines, cuboids, and rotated bounding boxes serve as indispensable tools in computer vision, each tailored to address specific challenges and scenarios. Polylines enable intricate shape delineation, enhancing applications like image segmentation and contour tracking. Cuboids, or 3D bounding boxes, are invaluable for accurately capturing the spatial dimensions of objects, significantly improving object detection, tracking, and augmented reality tasks. Meanwhile, rotated bounding boxes cater to objects with non-standard orientations, greatly enhancing precision in tasks such as text detection and irregular object localization. These diverse labeling techniques empower computer vision professionals to more effectively and precisely address a wide array of real-world problems, facilitating better representation and understanding of complex visual data. There is no better place to implement these labels than FiftyOne!

Join the FiftyOne Community!

Join the thousands of engineers and data scientists already using FiftyOne to solve some of the most challenging problems in computer vision today!