Scene video and scanpath mapping#
In this tutorial, we will map gaze data from an eye-tracking recording to video frames, estimate a scanpath, and overlay the gaze fixations on the video. We will use the pyneon library to work with Neon eye-tracking recordings, which contain video and event data, including gaze information.
1. Setup: Loading a Neon Recording#
First, we load the Neon recording, which contains video and gaze data. Ensure that you have installed the required libraries such as pyneon and have the recording dataset available.
[1]:
# Import necessary libraries
import sys
import numpy as np
from pyneon import get_sample_data, Dataset, Recording
import pandas as pd
# Download sample data (if not existing) and return the path
sample_dir = get_sample_data("Artworks")
print(sample_dir)
dataset_dir = sample_dir / "Timeseries Data + Scene Video"
dataset = Dataset(dataset_dir)
print(dataset)
recording = dataset[0]
print(recording)
C:\Users\jan-gabriel.hartel\Documents\GitHub\PyNeon\data\Artworks
Dataset | 1 recordings
Recording ID: 9a141750-95ca-48ee-9693-53bbb896b87e
Wearer ID: c4f68887-e96c-467f-a901-0fc9fce09c0a
Wearer name: JGH
Recording start time: 2025-06-16 12:49:27.817000
Recording duration: 357.538s
exist filename path
3d_eye_states True 3d_eye_states.csv C:\Users\jan-gabriel.hartel\Documents\GitHub\PyNeon\data\Artworks\Timeseries Data + Scene Video\artworks-9a141750\3d_eye_states.csv
blinks True blinks.csv C:\Users\jan-gabriel.hartel\Documents\GitHub\PyNeon\data\Artworks\Timeseries Data + Scene Video\artworks-9a141750\blinks.csv
events True events.csv C:\Users\jan-gabriel.hartel\Documents\GitHub\PyNeon\data\Artworks\Timeseries Data + Scene Video\artworks-9a141750\events.csv
fixations True fixations.csv C:\Users\jan-gabriel.hartel\Documents\GitHub\PyNeon\data\Artworks\Timeseries Data + Scene Video\artworks-9a141750\fixations.csv
gaze True gaze.csv C:\Users\jan-gabriel.hartel\Documents\GitHub\PyNeon\data\Artworks\Timeseries Data + Scene Video\artworks-9a141750\gaze.csv
imu True imu.csv C:\Users\jan-gabriel.hartel\Documents\GitHub\PyNeon\data\Artworks\Timeseries Data + Scene Video\artworks-9a141750\imu.csv
labels True labels.csv C:\Users\jan-gabriel.hartel\Documents\GitHub\PyNeon\data\Artworks\Timeseries Data + Scene Video\artworks-9a141750\labels.csv
saccades True saccades.csv C:\Users\jan-gabriel.hartel\Documents\GitHub\PyNeon\data\Artworks\Timeseries Data + Scene Video\artworks-9a141750\saccades.csv
world_timestamps True world_timestamps.csv C:\Users\jan-gabriel.hartel\Documents\GitHub\PyNeon\data\Artworks\Timeseries Data + Scene Video\artworks-9a141750\world_timestamps.csv
scene_video_info True scene_camera.json C:\Users\jan-gabriel.hartel\Documents\GitHub\PyNeon\data\Artworks\Timeseries Data + Scene Video\artworks-9a141750\scene_camera.json
scene_video True 11f35cc2_0.0-357.538.mp4 C:\Users\jan-gabriel.hartel\Documents\GitHub\PyNeon\data\Artworks\Timeseries Data + Scene Video\artworks-9a141750\11f35cc2_0.0-357.538.mp4
2. Mapping Gaze Data to Video Frames#
In Neon recordings, gaze events are not naturally synchronized with the video. To map gaze data to specific video frames, we can use the map_gaze_to_video method. This method requires the pyneon.video object for determination of video timestamps, the pyneon.fixations object to make use of PupilLabs fixation detection pipeline and the pyneon.gaze object for improved time resolution of gaze estimation.
By default, Neon reports fixations with a single coordinate. This is computed as average between all gaze coordinates over the interval dennoted as a fixation. However, this clashes with the funcional definition of a fixation as tracking a fixed point in space, used by Neon.
Imagine looking at a fixed point, for example a street sign, while you are walking past it. Despite the movement of your body and the relative movement of the sign, the fixation will be stabilised. As such, taking an average gaze coordinate over the enntire duration will not correspond to the location of the sign, or the fixation, ar any given point in time. Feeding this point into an optical flow algorithm would, with high likelihood, lead to tracking anything but the sign.
Therefore, we use partial averages of gaze locations around the respective frame’s timestamp. As the video is sampled at 30Hz while the gaze output nominally reaches 200Hz, we expect to take the average over 6 subsequent gaze points. This achieves a trade-off between recency of the reported gaze position at the given frame and error minimisation, by averaging over microsaccades around the actual fixation target as well as random errors.
[2]:
# Map gaze data to the video timestamps
synced_gaze = recording.sync_gaze_to_video(overwrite=True)
# Inspect the mapped gaze data
print(synced_gaze.data.tail())
gaze x [px] gaze y [px] worn fixation id blink id \
timestamp [ns]
1750071325151422222 NaN NaN <NA> <NA> <NA>
1750071325201422222 NaN NaN <NA> <NA> <NA>
1750071325251422222 NaN NaN <NA> <NA> <NA>
1750071325301422222 NaN NaN <NA> <NA> <NA>
1750071325351422222 NaN NaN <NA> <NA> <NA>
azimuth [deg] elevation [deg] frame_idx
timestamp [ns]
1750071325151422222 NaN NaN 10690
1750071325201422222 NaN NaN 10691
1750071325251422222 NaN NaN 10692
1750071325301422222 NaN NaN 10693
1750071325351422222 NaN NaN 10694
Above, we can see that each frame gets a current gaze position as well as a fixation status. Currently, three types of fixation status are used:
startdenoting the first frame corresponding to a fixationduringcorresponding to intermediate frames of the same fixationenddenoting the last frame of the fixation
This determination will become relevant for tracking the scanpath with optical flow. After all, while a fixation is still active, we get up-to-date gaze information. Only after its end, tracking becomes necessary.
3. Estimating the Scanpath#
Having matched every frame with a gaze coordinate, we can now get into the meat of the scanpath estimation. In dynamic scenes, the same object will not occupy the same scene-camera location over time. Therefore, we need to continuously map past fixation points as long as they are still visible in the frame.
The estimate_scanpath method achieves this by feeding fixation point denoted as end into a Lucas-Kanade sparse optical flow algorithm. This algorithm compares the video in vicinity of the point with the subsequent frame, updating the location in dependence of its movement. While a point is tracked, its status is flagged as tracked. In practice, many scene frames will have multiple simultaneously present past fixations. Our implementation carries them and repeately performs an optical
flow estimation for each point. Only when they can no longer be tracked, will they be flagged as lost and subsequently dropped for the next frame.
It should be noted that this algorithm is not optimised for performance and that it will take a considerable amount of time to run on limited hardware. For our computers, the algorithm takes roughly half the time of the video, though this benchmark heavily depends on the density of past fixation points and computational ressources
[3]:
# Estimate the scanpath based on the mapped gaze data
scanpath = recording.estimate_scanpath()
# Inspect the estimated scanpath
print(scanpath.data.head())
Loading saved scanpath from C:\Users\jan-gabriel.hartel\Documents\GitHub\PyNeon\data\Artworks\Timeseries Data + Scene Video\artworks-9a141750\derivatives\scanpath.pkl
fixations \
timestamp [ns]
1750070967817000000 fixation id gaze x [px] gaze y [px] fixation...
1750070967867000000 fixation id gaze x [px] gaze y [px] fixation...
1750070967917000000 fixation id gaze x [px] gaze y [px] fixation...
1750070967967000000 fixation id gaze x [px] gaze y [px] fixation...
1750070968017000000 fixation id gaze x [px] gaze y [px] fixation...
frame_idx
timestamp [ns]
1750070967817000000 0
1750070967867000000 1
1750070967917000000 2
1750070967967000000 3
1750070968017000000 4
We should take a moment to understand the format of the scanpath.data. As we care about getting a scanpath mapped on every single video-frame, we create it as a dataframe of dataframes. As such, every row carries both the timestamp as well as the frame index of the underlying video and saves a dataframe in the fixations cell. In this dataframe, every present fixation is provided with an id, coordinates and a fixation status, as seen below. The benefit of treating is a dataframe is the
possibility to use intuitive pandas indexing, allowing us, for example, to get a list of fixations at frame 2000.
As a quirk of Neon taking some time to start up, the first frames will usually not yield any usable results. Still, we carry them for consistency.
[4]:
# print fixations when column frame_idx is 1334. Frame_idx is not the idx of the dataframe, but the index of the video frame.
print(scanpath.data.loc[scanpath.data["frame_idx"] == 2000, "fixations"].values[0])
fixation id gaze x [px] gaze y [px] fixation status
0 174 903.107571 428.153429 during
1 173 924.962036 437.257629 tracked
2 172 922.353638 493.392975 tracked
3 171 960.446289 465.010101 tracked
4 170 835.920654 440.765076 tracked
5 169 833.580444 467.248718 tracked
6 168 698.10437 662.543335 tracked
7 167 693.992065 630.123047 tracked
8 166 713.760315 561.713867 tracked
9 165 543.353333 473.653381 tracked
10 164 698.986816 486.05957 tracked
11 163 704.265442 439.653992 tracked
12 162 750.842773 539.532349 tracked
4. Understanding Fixation Status#
Each fixation is assigned a status that indicates its lifecycle:
start: first frame of fixation
during: intermediate frames of fixation
end: last frame of fixation
tracked: Optical flow algorithm tracks fixation
lost: Tracking is lost, fixation is no longer tracked and gets dropped
5. Overlaying Fixations on the Video#
Now that we have the scanpath, we can overlay the gaze fixations on the video. This creates a video output with overlaid fixations, where:
A blue dot represents the current gaze location.
Green dots represent tracked fixations.
A red dot indicates no fixation (saccades or blinks).
Further, we draw connecting lines between past fixations to show the scanpath for the current video. The show_video option creates a live-output of the video rendering, but also increases the runtime.
[5]:
# Overlay the scanpath on the video and show the output
recording.overlay_scanpath(show_video=True)
Loading saved scanpath from C:\Users\jan-gabriel.hartel\Documents\GitHub\PyNeon\data\Artworks\Timeseries Data + Scene Video\artworks-9a141750\derivatives\scanpath.pkl
Overlay video already exists at C:\Users\jan-gabriel.hartel\Documents\GitHub\PyNeon\data\Artworks\Timeseries Data + Scene Video\artworks-9a141750\derivatives\scanpath.mp4; skipping render.
`show_video=True` has no effect because rendering was skipped.
Summary#
Mapping Gaze to Video: We used the
map_gaze_to_videomethod to match gaze data with video frames based on timestamps.Estimating Scanpath: The scanpath was estimated using
estimate_scanpath, which tracks fixations and uses optical flow to follow past fixations across scene changes.Overlaying Fixations: The fixations were visualized on the video by calling
overlay_fixations_on_video.
This workflow can be used to process eye-tracking data, align it with video frames, and visualize gaze movements within video recordings.