Skip to content

Commit

Permalink
Merge pull request #727 from pupil-labs/fix-invisible-docs
Browse files Browse the repository at this point in the history 
Fix invisible docs
  • Loading branch information
@marc-tonsen
marc-tonsen authored Dec 20, 2024
2 parents 4058267 + 44ab2b8 commit 47be814
Show file tree
Hide file tree
Showing 2 changed files with 204 additions and 3 deletions.
2 changes: 0 additions & 2 deletions invisible/data-collection/data-streams/index.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -22,8 +22,6 @@ Gaze data is output in pixel space of the scene camera image, which has a resolu
The gaze estimation algorithm is based on end-2-end deep learning and provides gaze data robustly without requiring a calibration. You can find a high-level description as well as a thorough evaluation of the accuracy and robustness of the algorithm in our [white paper](https://arxiv.org/pdf/2009.00508).

## Fixations & Saccades

<Badge>Pupil Cloud</Badge><Badge>Neon Player</Badge>
The two primary types of eye movements exhibited by the visual system are fixations and saccades. During fixations, the eyes are directed at a specific point in the environment. A saccade is a very quick movement where the eyes jump from one fixation to the next. Properties like the fixation duration are of significant importance for studying gaze behavior.

![Fixations](./fixations.jpg)
Expand Down
205 changes: 204 additions & 1 deletion invisible/pupil-cloud/enrichments/reference-image-mapper/index.md
View file
Original file line number Diff line number Diff line change
@@ -1 +1,204 @@
<!--@include: ../../../../neon/pupil-cloud/enrichments/reference-image-mapper/index.md-->
# Reference Image Mapper

Reference Image Mapper is a powerful tool to automatically map gaze onto features/objects in the environment.

<Youtube src="IF8f1Z3ZkEo" muted="1"/>

A heatmap of gaze data mapped onto the reference image can be generated, and mapped gaze and fixation data can further be downloaded as [CSV files](/pupil-cloud/enrichments/reference-image-mapper/#export-format).

## Setup

<Youtube src="ygqzQEzUIS4"/>

As described in the setup video, you will need two things in addition to your eye tracking recording(s) to produce a Reference Image Mapper enrichment:

1. A reference image
2. A scanning video of the object/feature(s) taken with the scene camera of Pupil Invisible

:::tip
**Reference Image**<br/>
Only the scanning recording needs to be taken with Pupil Invisible. The reference image can be taken with any camera. It can be also be a scanned image.
<br/>
<br/>
<b>Duration</b><br/>
The **3 min** duration limit only applies to the scanning recording, the rest of your videos can be as long as you like. <br/>But to avoid long processing times and get the best results, we recommend spliting the recordings into relevant parts where you expect the image to be mapped. <br/>
:::

In this guide, we will show you how the Reference Image Mapper can be used in a few different situations. This should help you get set up using the enrichment in your own testing environment!

Below, gaze is mapped in four very different environments: to a **magazine cover**, a **basketball backboard**, a **supermarket shelf**, and even a **whole building**!

![Reference images collage](./rim-collage.webp)

Let's take a look at what the Reference Image and Scanning Recording look like to produce these heatmaps.

### 1. Magazine Reading

<div class="grid grid-cols-1 sm:grid-cols-2 gap-4">
<div class="grid subgrid" style="grid-row: 1 / 4">
<h4>Reference image</h4>
<img src="./magazine-img.jpg" alt="Magazine Reference Image">
<div>First, we need a high-resolution .jpeg of the page.</div>
</div>
<div class="grid subgrid" style="grid-row: 1 / 4">
<h4>Scanning video</h4>
<div class="m-auto">
<Youtube src="SplYZK-ZzE8"/>
</div>
<div>
Then, we need a scanning recording, about <i>15 s</i> long, taken on a blank background and in good lighting (natural light works well). Note that the magazine page is clear with good contrast
</div>
</div>
</div>

### 2. Basketball

<div class="grid grid-cols-2 gap-4">
<div class="grid subgrid" style="grid-row: 1 / 4">
<h4>Reference image</h4>
<img src="./basketball-img.jpg" alt="Basketball Reference Image">
<div>Here we can take a photo of the basketball hoop and court background.</div>
</div>
<div class="grid subgrid" style="grid-row: 1 / 4">
<h4>Scanning video</h4>
<div class="m-auto">
<Youtube src="4MB5o4W_XJo"/>
</div>
<div>
The scanning recording, in this case, is about 45 s long – note that the angles and distances cover what a player might see when dribbling towards the basket or taking a shot
</div>
</div>
</div>

### 3. Supermarket Shelf

<div class="grid grid-cols-2 gap-4">
<div class="grid subgrid" style="grid-row: 1 / 4">
<h4>Reference image</h4>
<img src="./supermarket-img.jpg" alt="Supermarket Reference Image">
<div>This photo captures the assortment of packagíng in the coffee aisle of a supermarket.</div>
</div>
<div class="grid subgrid" style="grid-row: 1 / 4">
<h4>Scanning video</h4>
<div class="m-auto">
<Youtube src="d3Yk3nKDIOQ"/>
</div>
<div>
Notice that the scanning recording is fairly slow to reduce motion blur for this feature-rich shelf.
</div>
</div>
</div>

### 4. An Entire Building

<div class="grid grid-cols-2 gap-4">
<div class="grid subgrid" style="grid-row: 1 / 4">
<h4>Reference image</h4>
<img src="./building-img.jpg" alt="Building Reference Image">
<div>This is a photo of the <i>entire</i> building</div>
</div>
<div class="grid subgrid" style="grid-row: 1 / 4">
<h4>Scanning video</h4>
<div class="m-auto">
<Youtube src="0U4H-uOIHlw"/>
</div>
<div>
We take a longer scanning recording, about 2 min. The angles and distances cover what a person might see whilst walking past or standing in front of the building.
</div>
</div>
</div>

## Scanning Best Practices

Try to follow these best practices when recording a scanning video:

- Make the recording while holding Pupil Invisible in your hand rather than wearing it on your head.
- Record the object of interest from all possible angles and from all distances a subject may view them. More diversity is better. Collecting sufficiently diverse viewpoints for a large object, like a building, may require you to move accordingly large distances.
- Move the glasses slowly while recording to avoid motion blur.
- Make sure to have good contrast and that your scene lighting during scanning is similar to that during mapping.

What works and what doesn’t?

- The scene must be **feature-rich** like in the examples above. There has to be enough salient visual content to produce a successful mapping.
- The scene needs to have relatively **static** features in the environment. If there is a lot of movement or the objects change in appearance or shape, the mapping _can_ fail.

::: tip
**Ready to go?**<br>
Why not try replicating the above examples? Or even try it with your own use cases! If you haven’t already, we recommend you check out the [Cloud Getting Started Guide](/pupil-cloud/tutorials/), which covers the basics of working with enrichments.
:::

## Validate the Gaze Mapping

<video width="100%" controls>
<source src="./validate_rim.mp4" type="video/mp4">
</video>

To check if gaze has been mapped successfully, use the side-by-side view:

1. Select a recording.
2. Select the Reference Image Mapper Enrichment.
3. Select the Scene / Reference Image View.
<!-- 4. If you want to visualize and evaluate the 3D model generated (white dots), just turn on the Point Cloud toggle! Not in the new cloud UI (to be deleted if no toggle is added)-->

Now when you play back the recording you can see where gaze is mapped to on your reference image for validation.

## Occlusions

![Basketball Occlusion](./basketball-occlusion.png)

Sometimes an object will occlude the feature/object of interest. The reference image mapper may not recognize this, so a false positive mapping could occur.

If you need to remove the falsely mapped data points, there are a few workarounds.

1. Use Events and Sections to trim out false positive sections.
2. Manually remove the affected data points in the reference image mapper export, by finding the timestamp and deleting the row in the .csv export.

## Repetitions

In cases such as supermarket shopping, where features of the environment like freezers and aisles are repetitive and overlapping, it can be useful to divide recordings into shorter sections using [Events](/data-collection/events/) for enrichment. This way you can ensure gaze is only mapped to portions of the recording during which you know the user is looking at a particular part of the store.

::: tip
**Want to know more?**<br>
Under the hood, the Reference Image Mapper uses a method called Structure from Motion (SfM) to build a model of the environment. You can see this model as a ‘point cloud’ on the left side video.
:::

## Export Format

### gaze.csv

This file contains all the mapped gaze data from all sections.

| Field | Description |
| ------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| **section id** | Unique identifier of the corresponding section. |
| **recording id** | Unique identifier of the recording this sample belongs to. |
| **timestamp [ns]** | UTC timestamp in nanoseconds of the sample. Equal to the timestamp of the original gaze sample before mapping. |
| **gaze detected in reference image** | Boolean indicating whether the gaze point was detected inside or outside of the reference image. |
| **gaze position in reference image x [px]** | Float value representing the x-coordinate of the mapped gaze point in pixel coordinates. If the reference image was not detected in the scene at the given time this value is empty. |
| **gaze position in reference image y [px]** | Same as "gaze position in reference image x [px]" but for the y-coordinate. |
| **fixation id** | If this gaze sample belongs to a fixation event, this is the corresponding id of the fixation. Otherwise, this field is empty. |
| **blink id** | If this gaze samples belongs to a blink event, this is the corresponding id of the blink. Otherwise this field is empty. |

::: info
This CSV file only contains data-points where the reference image has been localised in the scene. Looking for all the gaze points? Check [this file.](/data-collection/data-format/#gaze-csv)
:::

### fixations.csv

This file contains fixation events detected in the gaze data stream and mapped to the reference image.

| Field | Description |
| ---------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **section id** | Unique identifier of the corresponding section. |
| **recording id** | Unique identifier of the recording this sample belongs to. |
| **fixation id** | Identifier of fixation within the section. The id corresponds to the fixation id of the raw unmapped data. |
| **start&nbsp;timestamp&nbsp;[ns]** | UTC timestamp in nanoseconds of the start of the fixation. |
| **end&nbsp;timestamp&nbsp;[ns]** | UTC timestamp in nanoseconds of the end of the fixation. |
| **duration&nbsp;[ms]** | Duration of the fixation in milliseconds. |
| **fixation detected in reference image** | Boolean indicating whether the fixation was inside or outside of the reference image. |
| **fixation x [px]** | Float value representing the x-coordinate of the fixation in reference image coordinates. This position is the average of all mapped gaze samples within the fixation. |
| **fixation y [px]** | Same as "fixation x [px]" but for the y-coordinate. |

### Reference Image

The reference image that was used for defining the enrichment. The file is named `reference_image.jpeg|png`.

0 comments on commit 47be814

Please sign in to comment.